diff --git a/EMG_Arm/src/CMakeLists.txt b/EMG_Arm/src/CMakeLists.txt
index 3ddece0..9a7dd7d 100644
--- a/EMG_Arm/src/CMakeLists.txt
+++ b/EMG_Arm/src/CMakeLists.txt
@@ -20,6 +20,7 @@ set(DRIVER_SOURCES
 
 set(CORE_SOURCES
     core/gestures.c
+    core/inference.c
 )
 
 set(APP_SOURCES
diff --git a/EMG_Arm/src/app/main.c b/EMG_Arm/src/app/main.c
index e16d68b..38db647 100644
--- a/EMG_Arm/src/app/main.c
+++ b/EMG_Arm/src/app/main.c
@@ -12,17 +12,18 @@
  * @note This is Layer 4 (Application).
  */
 
+#include "esp_timer.h"
+#include <freertos/FreeRTOS.h>
+#include <freertos/queue.h>
+#include <freertos/task.h>
 #include <stdio.h>
 #include <string.h>
-#include <freertos/FreeRTOS.h>
-#include <freertos/task.h>
-#include <freertos/queue.h>
-#include "esp_timer.h"
 
 #include "config/config.h"
-#include "drivers/hand.h"
-#include "drivers/emg_sensor.h"
 #include "core/gestures.h"
+#include "core/inference.h" // [NEW]
+#include "drivers/emg_sensor.h"
+#include "drivers/hand.h"
 
 /*******************************************************************************
  * Constants
@@ -39,20 +40,22 @@
  * @brief Device state machine.
  */
 typedef enum {
-    STATE_IDLE = 0,       /**< Waiting for connect command */
-    STATE_CONNECTED,      /**< Connected, waiting for start command */
-    STATE_STREAMING,      /**< Actively streaming EMG data */
+  STATE_IDLE = 0,   /**< Waiting for connect command */
+  STATE_CONNECTED,  /**< Connected, waiting for start command */
+  STATE_STREAMING,  /**< Actively streaming raw EMG data (for training) */
+  STATE_PREDICTING, /**< [NEW] On-device inference and control */
 } device_state_t;
 
 /**
  * @brief Commands from host.
  */
 typedef enum {
-    CMD_NONE = 0,
-    CMD_CONNECT,
-    CMD_START,
-    CMD_STOP,
-    CMD_DISCONNECT,
+  CMD_NONE = 0,
+  CMD_CONNECT,
+  CMD_START,         /**< Start raw streaming */
+  CMD_START_PREDICT, /**< [NEW] Start on-device prediction */
+  CMD_STOP,
+  CMD_DISCONNECT,
 } command_t;
 
 /*******************************************************************************
@@ -80,38 +83,39 @@ static void send_ack_connect(void);
  * @param line Input line from serial
  * @return Parsed command
  */
-static command_t parse_command(const char* line)
-{
-    /* Simple JSON parsing - look for "cmd" field */
-    const char* cmd_start = strstr(line, "\"cmd\"");
-    if (!cmd_start) {
-        return CMD_NONE;
-    }
-
-    /* Find the value after "cmd": */
-    const char* value_start = strchr(cmd_start, ':');
-    if (!value_start) {
-        return CMD_NONE;
-    }
-
-    /* Skip whitespace and opening quote */
-    value_start++;
-    while (*value_start == ' ' || *value_start == '"') {
-        value_start++;
-    }
-
-    /* Match command strings */
-    if (strncmp(value_start, "connect", 7) == 0) {
-        return CMD_CONNECT;
-    } else if (strncmp(value_start, "start", 5) == 0) {
-        return CMD_START;
-    } else if (strncmp(value_start, "stop", 4) == 0) {
-        return CMD_STOP;
-    } else if (strncmp(value_start, "disconnect", 10) == 0) {
-        return CMD_DISCONNECT;
-    }
-
+static command_t parse_command(const char *line) {
+  /* Simple JSON parsing - look for "cmd" field */
+  const char *cmd_start = strstr(line, "\"cmd\"");
+  if (!cmd_start) {
     return CMD_NONE;
+  }
+
+  /* Find the value after "cmd": */
+  const char *value_start = strchr(cmd_start, ':');
+  if (!value_start) {
+    return CMD_NONE;
+  }
+
+  /* Skip whitespace and opening quote */
+  value_start++;
+  while (*value_start == ' ' || *value_start == '"') {
+    value_start++;
+  }
+
+  /* Match command strings */
+  if (strncmp(value_start, "connect", 7) == 0) {
+    return CMD_CONNECT;
+  } else if (strncmp(value_start, "start_predict", 13) == 0) {
+    return CMD_START_PREDICT;
+  } else if (strncmp(value_start, "start", 5) == 0) {
+    return CMD_START;
+  } else if (strncmp(value_start, "stop", 4) == 0) {
+    return CMD_STOP;
+  } else if (strncmp(value_start, "disconnect", 10) == 0) {
+    return CMD_DISCONNECT;
+  }
+
+  return CMD_NONE;
 }
 
 /*******************************************************************************
@@ -120,224 +124,213 @@ static command_t parse_command(const char* line)
 
 /**
  * @brief FreeRTOS task to read serial input and parse commands.
- *
- * This task runs continuously, reading lines from stdin (USB serial)
- * and updating device state directly. This allows commands to interrupt
- * streaming immediately via the volatile state variable.
- *
- * @param pvParameters Unused
  */
-static void serial_input_task(void* pvParameters)
-{
-    char line_buffer[CMD_BUFFER_SIZE];
-    int line_idx = 0;
+static void serial_input_task(void *pvParameters) {
+  char line_buffer[CMD_BUFFER_SIZE];
+  int line_idx = 0;
 
-    while (1) {
-        /* Read one character at a time */
-        int c = getchar();
+  while (1) {
+    int c = getchar();
 
-        if (c == EOF || c == 0xFF) {
-            /* No data available, yield to other tasks */
-            vTaskDelay(pdMS_TO_TICKS(10));
-            continue;
-        }
-
-        if (c == '\n' || c == '\r') {
-            /* End of line - process command */
-            if (line_idx > 0) {
-                line_buffer[line_idx] = '\0';
-
-                command_t cmd = parse_command(line_buffer);
-
-                if (cmd != CMD_NONE) {
-                    /* Handle CONNECT command from ANY state (for reconnection/recovery) */
-                    if (cmd == CMD_CONNECT) {
-                        /* Stop streaming if active, reset to CONNECTED state */
-                        g_device_state = STATE_CONNECTED;
-                        send_ack_connect();
-                        printf("[STATE] ANY -> CONNECTED (reconnect)\n");
-                    }
-                    /* Handle other state transitions */
-                    else {
-                        switch (g_device_state) {
-                            case STATE_IDLE:
-                                /* Only CONNECT allowed from IDLE (handled above) */
-                                break;
-
-                            case STATE_CONNECTED:
-                                if (cmd == CMD_START) {
-                                    g_device_state = STATE_STREAMING;
-                                    printf("[STATE] CONNECTED -> STREAMING\n");
-                                    /* Signal state machine to start streaming */
-                                    xQueueSend(g_cmd_queue, &cmd, 0);
-                                } else if (cmd == CMD_DISCONNECT) {
-                                    g_device_state = STATE_IDLE;
-                                    printf("[STATE] CONNECTED -> IDLE\n");
-                                }
-                                break;
-
-                            case STATE_STREAMING:
-                                if (cmd == CMD_STOP) {
-                                    g_device_state = STATE_CONNECTED;
-                                    printf("[STATE] STREAMING -> CONNECTED\n");
-                                    /* Streaming loop will exit when it sees state change */
-                                } else if (cmd == CMD_DISCONNECT) {
-                                    g_device_state = STATE_IDLE;
-                                    printf("[STATE] STREAMING -> IDLE\n");
-                                    /* Streaming loop will exit when it sees state change */
-                                }
-                                break;
-                        }
-                    }
-                }
-
-                line_idx = 0;
-            }
-        } else if (line_idx < CMD_BUFFER_SIZE - 1) {
-            /* Add character to buffer */
-            line_buffer[line_idx++] = (char)c;
-        } else {
-            /* Buffer overflow - reset */
-            line_idx = 0;
-        }
+    if (c == EOF || c == 0xFF) {
+      vTaskDelay(pdMS_TO_TICKS(10));
+      continue;
     }
+
+    if (c == '\n' || c == '\r') {
+      if (line_idx > 0) {
+        line_buffer[line_idx] = '\0';
+        command_t cmd = parse_command(line_buffer);
+
+        if (cmd != CMD_NONE) {
+          if (cmd == CMD_CONNECT) {
+            g_device_state = STATE_CONNECTED;
+            send_ack_connect();
+            printf("[STATE] ANY -> CONNECTED (reconnect)\n");
+          } else {
+            switch (g_device_state) {
+            case STATE_IDLE:
+              break;
+
+            case STATE_CONNECTED:
+              if (cmd == CMD_START) {
+                g_device_state = STATE_STREAMING;
+                printf("[STATE] CONNECTED -> STREAMING\n");
+                xQueueSend(g_cmd_queue, &cmd, 0);
+              } else if (cmd == CMD_START_PREDICT) {
+                g_device_state = STATE_PREDICTING;
+                printf("[STATE] CONNECTED -> PREDICTING\n");
+                xQueueSend(g_cmd_queue, &cmd, 0);
+              } else if (cmd == CMD_DISCONNECT) {
+                g_device_state = STATE_IDLE;
+                printf("[STATE] CONNECTED -> IDLE\n");
+              }
+              break;
+
+            case STATE_STREAMING:
+            case STATE_PREDICTING:
+              if (cmd == CMD_STOP) {
+                g_device_state = STATE_CONNECTED;
+                printf("[STATE] ACTIVE -> CONNECTED\n");
+              } else if (cmd == CMD_DISCONNECT) {
+                g_device_state = STATE_IDLE;
+                printf("[STATE] ACTIVE -> IDLE\n");
+              }
+              break;
+            }
+          }
+        }
+        line_idx = 0;
+      }
+    } else if (line_idx < CMD_BUFFER_SIZE - 1) {
+      line_buffer[line_idx++] = (char)c;
+    } else {
+      line_idx = 0;
+    }
+  }
 }
 
 /*******************************************************************************
  * State Machine
  ******************************************************************************/
 
-/**
- * @brief Send JSON acknowledgment for connection.
- */
-static void send_ack_connect(void)
-{
-    printf("{\"status\":\"ack_connect\",\"device\":\"ESP32-EMG\",\"channels\":%d}\n",
-           EMG_NUM_CHANNELS);
-    fflush(stdout);
+static void send_ack_connect(void) {
+  printf(
+      "{\"status\":\"ack_connect\",\"device\":\"ESP32-EMG\",\"channels\":%d}\n",
+      EMG_NUM_CHANNELS);
+  fflush(stdout);
 }
 
 /**
- * @brief Stream EMG data continuously until stopped.
- *
- * This function blocks and streams data at the configured sample rate.
- * Returns when state changes from STREAMING.
+ * @brief Stream raw EMG data (Training Mode).
  */
-static void stream_emg_data(void)
-{
-    emg_sample_t sample;
-    const TickType_t delay_ticks = 1;  /* 1 tick = 1ms at 1000 Hz tick rate */
+static void stream_emg_data(void) {
+  emg_sample_t sample;
+  const TickType_t delay_ticks = 1;
 
-    while (g_device_state == STATE_STREAMING) {
-        /* Read EMG (fake or real depending on FEATURE_FAKE_EMG) */
-        emg_sensor_read(&sample);
-
-        /* Output in CSV format - channels only, Python handles timestamps */
-        printf("%u,%u,%u,%u\n",
-               sample.channels[0],
-               sample.channels[1],
-               sample.channels[2],
-               sample.channels[3]);
-
-        /* Yield to FreeRTOS scheduler - prevents watchdog timeout */
-        vTaskDelay(delay_ticks);
-    }
-}
-
-/**
- * @brief Main state machine loop.
- *
- * Monitors device state and starts streaming when requested.
- * Serial input task handles all state transitions directly.
- */
-static void state_machine_loop(void)
-{
-    command_t cmd;
-    const TickType_t poll_interval = pdMS_TO_TICKS(50);
-
-    while (1) {
-        /* Check if we should start streaming */
-        if (g_device_state == STATE_STREAMING) {
-            /* Stream until state changes (via serial input task) */
-            stream_emg_data();
-            /* Returns when state is no longer STREAMING */
-        }
-
-        /* Wait for start command or just poll state */
-        /* Timeout allows checking state even if queue is empty */
-        xQueueReceive(g_cmd_queue, &cmd, poll_interval);
-
-        /* Note: State transitions are handled by serial_input_task */
-        /* This loop only triggers streaming when state becomes STREAMING */
-    }
-}
-
-void emgPrinter() {
-  TickType_t previousWake = xTaskGetTickCount();
-  while (1) {
-    emg_sample_t sample;
+  while (g_device_state == STATE_STREAMING) {
     emg_sensor_read(&sample);
-    for (uint8_t i = 0; i < EMG_NUM_CHANNELS; i++) {
-      printf("%d", sample.channels[i]);
-      if (i != EMG_NUM_CHANNELS - 1) printf(" | ");
+    printf("%u,%u,%u,%u\n", sample.channels[0], sample.channels[1],
+           sample.channels[2], sample.channels[3]);
+    vTaskDelay(delay_ticks);
+  }
+}
+
+/**
+ * @brief Run on-device inference (Prediction Mode).
+ */
+static void run_inference_loop(void) {
+  emg_sample_t sample;
+  const TickType_t delay_ticks = 1; // 1ms @ 1kHz
+  int last_gesture = -1;
+
+  // Reset inference state
+  inference_init();
+  printf("{\"status\":\"info\",\"msg\":\"Inference started\"}\n");
+
+  while (g_device_state == STATE_PREDICTING) {
+    emg_sensor_read(&sample);
+
+    // Add to buffer
+    // Note: sample.channels is uint16_t, matching inference engine expectation
+    if (inference_add_sample(sample.channels)) {
+      // Buffer full (sliding window), run prediction
+      // We can optimize stride here (e.g. valid prediction only every N
+      // samples) For now, let's predict every sample (sliding window) or
+      // throttle if too slow. ESP32S3 is fast enough for 4ch features @ 1kHz?
+      // maybe. Let's degrade to 50Hz updates (20ms stride) to be safe and avoid
+      // UART spam.
+
+      static int stride_counter = 0;
+      stride_counter++;
+
+      if (stride_counter >= 20) { // 20ms stride
+        float confidence = 0;
+        int gesture_idx = inference_predict(&confidence);
+        stride_counter = 0;
+
+        if (gesture_idx >= 0) {
+          // Map class index (0-N) to gesture enum (correct hardware action)
+          int gesture_enum = inference_get_gesture_enum(gesture_idx);
+
+          // Execute gesture on hand
+          gestures_execute((gesture_t)gesture_enum);
+
+          // Send telemetry if changed or periodically?
+          // "Live prediction flow should change to only have each new output...
+          // sent"
+          if (gesture_idx != last_gesture) {
+            printf("{\"gesture\":\"%s\",\"conf\":%.2f}\n",
+                   inference_get_class_name(gesture_idx), confidence);
+            last_gesture = gesture_idx;
+          }
+        }
+      }
     }
-    printf("\n");
-    // vTaskDelayUntil(&previousWake, pdMS_TO_TICKS(100));
+
+    vTaskDelay(delay_ticks);
+  }
+}
+
+static void state_machine_loop(void) {
+  command_t cmd;
+  const TickType_t poll_interval = pdMS_TO_TICKS(50);
+
+  while (1) {
+    if (g_device_state == STATE_STREAMING) {
+      stream_emg_data();
+    } else if (g_device_state == STATE_PREDICTING) {
+      run_inference_loop();
+    }
+
+    xQueueReceive(g_cmd_queue, &cmd, poll_interval);
   }
 }
 
 void appConnector() {
-  /* Create command queue */
-    g_cmd_queue = xQueueCreate(10, sizeof(command_t));
-    if (g_cmd_queue == NULL) {
-        printf("[ERROR] Failed to create command queue!\n");
-        return;
-    }
+  g_cmd_queue = xQueueCreate(10, sizeof(command_t));
+  if (g_cmd_queue == NULL) {
+    printf("[ERROR] Failed to create command queue!\n");
+    return;
+  }
 
-    /* Launch serial input task */
-    xTaskCreate(
-        serial_input_task,
-        "serial_input",
-        4096,              /* Stack size */
-        NULL,              /* Parameters */
-        5,                 /* Priority */
-        NULL               /* Task handle */
-    );
+  xTaskCreate(serial_input_task, "serial_input", 4096, NULL, 5, NULL);
 
-    printf("[PROTOCOL] Waiting for host to connect...\n");
-    printf("[PROTOCOL] Send: {\"cmd\": \"connect\"}\n\n");
+  printf("[PROTOCOL] Waiting for host to connect...\n");
+  printf("[PROTOCOL] Send: {\"cmd\": \"connect\"}\n");
+  printf("[PROTOCOL] Send: {\"cmd\": \"start_predict\"} for on-device "
+         "inference\n\n");
 
-    /* Run main state machine */
-    state_machine_loop();
+  state_machine_loop();
 }
 
 /*******************************************************************************
  * Application Entry Point
  ******************************************************************************/
 
-void app_main(void)
-{
-    printf("\n");
-    printf("========================================\n");
-    printf("  Bucky Arm - EMG Robotic Hand\n");
-    printf("  Firmware v2.0.0 (Handshake Protocol)\n");
-    printf("========================================\n\n");
+void app_main(void) {
+  printf("\n");
+  printf("========================================\n");
+  printf("  Bucky Arm - EMG Robotic Hand\n");
+  printf("  Firmware v2.1.0 (Inference Enabled)\n");
+  printf("========================================\n\n");
 
-    /* Initialize subsystems */
-    printf("[INIT] Initializing hand (servos)...\n");
-    hand_init();
+  printf("[INIT] Initializing hand (servos)...\n");
+  hand_init();
 
-    printf("[INIT] Initializing EMG sensor...\n");
-    emg_sensor_init();
+  printf("[INIT] Initializing EMG sensor...\n");
+  emg_sensor_init();
+
+  printf("[INIT] Initializing Inference Engine...\n");
+  inference_init();
 
 #if FEATURE_FAKE_EMG
-    printf("[INIT] Using FAKE EMG data (sensors not connected)\n");
+  printf("[INIT] Using FAKE EMG data (sensors not connected)\n");
 #else
-    printf("[INIT] Using REAL EMG sensors\n");
+  printf("[INIT] Using REAL EMG sensors\n");
 #endif
 
-    printf("[INIT] Done!\n\n");
+  printf("[INIT] Done!\n\n");
 
-    // emgPrinter();
-    appConnector();
+  appConnector();
 }
diff --git a/EMG_Arm/src/core/inference.c b/EMG_Arm/src/core/inference.c
new file mode 100644
index 0000000..4e800c3
--- /dev/null
+++ b/EMG_Arm/src/core/inference.c
@@ -0,0 +1,281 @@
+/**
+ * @file inference.c
+ * @brief Implementation of EMG inference engine.
+ */
+
+#include "inference.h"
+#include "config/config.h"
+#include "model_weights.h"
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+
+// --- Constants ---
+#define SMOOTHING_FACTOR 0.7f // EMA factor for probability (matches Python)
+#define VOTE_WINDOW 5         // Majority vote window size
+#define DEBOUNCE_COUNT 3      // Confirmations needed to change output
+
+// --- State ---
+static uint16_t window_buffer[INFERENCE_WINDOW_SIZE][NUM_CHANNELS];
+static int buffer_head = 0;
+static int samples_collected = 0;
+
+// Smoothing State
+static float smoothed_probs[MODEL_NUM_CLASSES];
+static int vote_history[VOTE_WINDOW];
+static int vote_head = 0;
+static int current_output = -1;
+static int pending_output = -1;
+static int pending_count = 0;
+
+void inference_init(void) {
+  memset(window_buffer, 0, sizeof(window_buffer));
+  buffer_head = 0;
+  samples_collected = 0;
+
+  // Initialize smoothing
+  for (int i = 0; i < MODEL_NUM_CLASSES; i++) {
+    smoothed_probs[i] = 1.0f / MODEL_NUM_CLASSES;
+  }
+  for (int i = 0; i < VOTE_WINDOW; i++) {
+    vote_history[i] = -1;
+  }
+  vote_head = 0;
+  current_output = -1;
+  pending_output = -1;
+  pending_count = 0;
+}
+
+bool inference_add_sample(uint16_t *channels) {
+  // Add to circular buffer
+  for (int i = 0; i < NUM_CHANNELS; i++) {
+    window_buffer[buffer_head][i] = channels[i];
+  }
+
+  buffer_head = (buffer_head + 1) % INFERENCE_WINDOW_SIZE;
+
+  if (samples_collected < INFERENCE_WINDOW_SIZE) {
+    samples_collected++;
+    return false;
+  }
+
+  return true; // Buffer is full (always ready in sliding window, but caller
+               // controls stride)
+}
+
+// --- Feature Extraction ---
+
+static void compute_features(float *features_out) {
+  // Process each channel
+  // We need to iterate over the logical window (unrolling circular buffer)
+
+  for (int ch = 0; ch < NUM_CHANNELS; ch++) {
+    float sum = 0;
+    float sq_sum = 0;
+
+    // Pass 1: Mean (for centering) and raw values collection
+    // We could optimize by not copying, but accessing logically is safer
+    float signal[INFERENCE_WINDOW_SIZE];
+
+    int idx = buffer_head; // Oldest sample
+    for (int i = 0; i < INFERENCE_WINDOW_SIZE; i++) {
+      signal[i] = (float)window_buffer[idx][ch];
+      sum += signal[i];
+      idx = (idx + 1) % INFERENCE_WINDOW_SIZE;
+    }
+
+    float mean = sum / INFERENCE_WINDOW_SIZE;
+
+    // Pass 2: Centering and Features
+    float wl = 0;
+    int zc = 0;
+    int ssc = 0;
+
+    // Center the signal
+    for (int i = 0; i < INFERENCE_WINDOW_SIZE; i++) {
+      signal[i] -= mean;
+      sq_sum += signal[i] * signal[i];
+    }
+
+    float rms = sqrtf(sq_sum / INFERENCE_WINDOW_SIZE);
+
+    // Thresholds
+    float zc_thresh = FEAT_ZC_THRESH * rms;
+    float ssc_thresh = (FEAT_SSC_THRESH * rms) *
+                       (FEAT_SSC_THRESH * rms); // threshold is on diff product
+
+    for (int i = 0; i < INFERENCE_WINDOW_SIZE - 1; i++) {
+      // WL
+      wl += fabsf(signal[i + 1] - signal[i]);
+
+      // ZC
+      if ((signal[i] > 0 && signal[i + 1] < 0) ||
+          (signal[i] < 0 && signal[i + 1] > 0)) {
+        if (fabsf(signal[i] - signal[i + 1]) > zc_thresh) {
+          zc++;
+        }
+      }
+
+      // SSC (needs 3 points, so loop to N-2)
+      if (i < INFERENCE_WINDOW_SIZE - 2) {
+        float diff1 = signal[i + 1] - signal[i];
+        float diff2 = signal[i + 1] - signal[i + 2];
+        if ((diff1 * diff2) > ssc_thresh) {
+          ssc++;
+        }
+      }
+    }
+
+    // Store features: [RMS, WL, ZC, SSC] per channel
+    int base = ch * 4;
+    features_out[base + 0] = rms;
+    features_out[base + 1] = wl;
+    features_out[base + 2] = (float)zc;
+    features_out[base + 3] = (float)ssc;
+  }
+}
+
+// --- Prediction ---
+
+int inference_predict(float *confidence) {
+  if (samples_collected < INFERENCE_WINDOW_SIZE) {
+    return -1;
+  }
+
+  // 1. Extract Features
+  float features[MODEL_NUM_FEATURES];
+  compute_features(features);
+
+  // 2. LDA Inference (Linear Score)
+  float raw_scores[MODEL_NUM_CLASSES];
+  float max_score = -1e9;
+  int max_idx = 0;
+
+  // Calculate raw discriminative scores
+  for (int c = 0; c < MODEL_NUM_CLASSES; c++) {
+    float score = LDA_INTERCEPTS[c];
+    for (int f = 0; f < MODEL_NUM_FEATURES; f++) {
+      score += features[f] * LDA_WEIGHTS[c][f];
+    }
+    raw_scores[c] = score;
+  }
+
+  // Convert scores to probabilities (Softmax)
+  // LDA scores are log-likelihoods + const. Softmax is appropriate.
+  float sum_exp = 0;
+  float probas[MODEL_NUM_CLASSES];
+
+  // Numerical stability: subtract max
+  // Create temp copy for max finding
+  for (int c = 0; c < MODEL_NUM_CLASSES; c++) {
+    if (raw_scores[c] > max_score)
+      max_score = raw_scores[c];
+  }
+
+  for (int c = 0; c < MODEL_NUM_CLASSES; c++) {
+    probas[c] = expf(raw_scores[c] - max_score);
+    sum_exp += probas[c];
+  }
+  for (int c = 0; c < MODEL_NUM_CLASSES; c++) {
+    probas[c] /= sum_exp;
+  }
+
+  // 3. Smoothing
+  // 3a. Probability EMA
+  float max_smoothed_prob = 0;
+  int smoothed_winner = 0;
+
+  for (int c = 0; c < MODEL_NUM_CLASSES; c++) {
+    smoothed_probs[c] = (SMOOTHING_FACTOR * smoothed_probs[c]) +
+                        ((1.0f - SMOOTHING_FACTOR) * probas[c]);
+
+    if (smoothed_probs[c] > max_smoothed_prob) {
+      max_smoothed_prob = smoothed_probs[c];
+      smoothed_winner = c;
+    }
+  }
+
+  // 3b. Majority Vote
+  vote_history[vote_head] = smoothed_winner;
+  vote_head = (vote_head + 1) % VOTE_WINDOW;
+
+  int counts[MODEL_NUM_CLASSES];
+  memset(counts, 0, sizeof(counts));
+
+  for (int i = 0; i < VOTE_WINDOW; i++) {
+    if (vote_history[i] != -1) {
+      counts[vote_history[i]]++;
+    }
+  }
+
+  int majority_winner = 0;
+  int majority_count = 0;
+  for (int c = 0; c < MODEL_NUM_CLASSES; c++) {
+    if (counts[c] > majority_count) {
+      majority_count = counts[c];
+      majority_winner = c;
+    }
+  }
+
+  // 3c. Debounce
+  int final_result = current_output;
+
+  if (current_output == -1) {
+    current_output = majority_winner;
+    pending_output = majority_winner;
+    pending_count = 1;
+    final_result = majority_winner;
+  } else if (majority_winner == current_output) {
+    pending_output = majority_winner;
+    pending_count = 1;
+  } else if (majority_winner == pending_output) {
+    pending_count++;
+    if (pending_count >= DEBOUNCE_COUNT) {
+      current_output = majority_winner;
+      final_result = majority_winner;
+    }
+  } else {
+    pending_output = majority_winner;
+    pending_count = 1;
+  }
+
+  // Use smoothed probability of the final winner as confidence
+  // Or simpler: use fraction of votes
+  *confidence = (float)majority_count / VOTE_WINDOW;
+
+  return final_result;
+}
+
+const char *inference_get_class_name(int class_idx) {
+  if (class_idx >= 0 && class_idx < MODEL_NUM_CLASSES) {
+    return MODEL_CLASS_NAMES[class_idx];
+  }
+  return "UNKNOWN";
+}
+
+int inference_get_gesture_enum(int class_idx) {
+  const char *name = inference_get_class_name(class_idx);
+
+  // Map string name to gesture_t enum
+  // Strings must match those in Python list: ["fist", "hook_em", "open",
+  // "rest", "thumbs_up"] Note: Python strings are lowercase, config.h enums
+  // are: GESTURE_NONE=0, REST=1, FIST=2, OPEN=3, HOOK_EM=4, THUMBS_UP=5
+
+  // Case-insensitive check would be safer, but let's assume Python output is
+  // lowercase as seen in scripts or uppercase if specified. In
+  // learning_data_collection.py, they seem to be "rest", "open", "fist", etc.
+
+  // Simple string matching
+  if (strcmp(name, "rest") == 0 || strcmp(name, "REST") == 0)
+    return GESTURE_REST;
+  if (strcmp(name, "fist") == 0 || strcmp(name, "FIST") == 0)
+    return GESTURE_FIST;
+  if (strcmp(name, "open") == 0 || strcmp(name, "OPEN") == 0)
+    return GESTURE_OPEN;
+  if (strcmp(name, "hook_em") == 0 || strcmp(name, "HOOK_EM") == 0)
+    return GESTURE_HOOK_EM;
+  if (strcmp(name, "thumbs_up") == 0 || strcmp(name, "THUMBS_UP") == 0)
+    return GESTURE_THUMBS_UP;
+
+  return GESTURE_NONE;
+}
diff --git a/EMG_Arm/src/core/inference.h b/EMG_Arm/src/core/inference.h
new file mode 100644
index 0000000..32cf3b0
--- /dev/null
+++ b/EMG_Arm/src/core/inference.h
@@ -0,0 +1,47 @@
+/**
+ * @file inference.h
+ * @brief On-device inference engine for EMG gesture recognition.
+ */
+
+#ifndef INFERENCE_H
+#define INFERENCE_H
+
+#include <stdbool.h>
+#include <stdint.h>
+
+// --- Configuration ---
+#define INFERENCE_WINDOW_SIZE 150 // Window size in samples (must match Python)
+#define NUM_CHANNELS 4            // Number of EMG channels
+
+/**
+ * @brief Initialize the inference engine.
+ */
+void inference_init(void);
+
+/**
+ * @brief Add a sample to the inference buffer.
+ *
+ * @param channels Array of 4 channel values (raw ADC)
+ * @return true if a full window is ready for processing
+ */
+bool inference_add_sample(uint16_t *channels);
+
+/**
+ * @brief Run inference on the current window.
+ *
+ * @param confidence Output pointer for confidence score (0.0 - 1.0)
+ * @return Detected class index (-1 if error)
+ */
+int inference_predict(float *confidence);
+
+/**
+ * @brief Get the name of a class index.
+ */
+const char *inference_get_class_name(int class_idx);
+
+/**
+ * @brief Map class index to gesture_t enum.
+ */
+int inference_get_gesture_enum(int class_idx);
+
+#endif /* INFERENCE_H */
diff --git a/EMG_Arm/src/core/model_weights.h b/EMG_Arm/src/core/model_weights.h
new file mode 100644
index 0000000..1dd5c94
--- /dev/null
+++ b/EMG_Arm/src/core/model_weights.h
@@ -0,0 +1,44 @@
+/**
+ * @file model_weights.h
+ * @brief Placeholder for trained model weights.
+ *
+ * This file should be generated by the Python script (Training Page -> Export for ESP32).
+ */
+
+#ifndef MODEL_WEIGHTS_H
+#define MODEL_WEIGHTS_H
+
+#include <stdint.h>
+
+/* Metadata */
+#define MODEL_NUM_CLASSES 5
+#define MODEL_NUM_FEATURES 16
+
+/* Class Names */
+static const char* MODEL_CLASS_NAMES[MODEL_NUM_CLASSES] = {
+    "REST",
+    "OPEN",
+    "FIST",
+    "HOOK_EM",
+    "THUMBS_UP",
+};
+
+/* Feature Extractor Parameters */
+#define FEAT_ZC_THRESH 0.1f
+#define FEAT_SSC_THRESH 0.1f
+
+/* LDA Intercepts/Biases */
+static const float LDA_INTERCEPTS[MODEL_NUM_CLASSES] = {
+    0.0f, 0.0f, 0.0f, 0.0f, 0.0f
+};
+
+/* LDA Coefficients (Weights) */
+static const float LDA_WEIGHTS[MODEL_NUM_CLASSES][MODEL_NUM_FEATURES] = {
+    {0.0f},
+    {0.0f},
+    {0.0f},
+    {0.0f},
+    {0.0f}
+};
+
+#endif /* MODEL_WEIGHTS_H */
diff --git a/emg_gui.py b/emg_gui.py
index 307d417..59bc103 100644
--- a/emg_gui.py
+++ b/emg_gui.py
@@ -1322,6 +1322,18 @@ class TrainingPage(BasePage):
         )
         self.train_button.pack(pady=20)
 
+        # Export button
+        self.export_button = ctk.CTkButton(
+            self.content,
+            text="Export for ESP32",
+            font=ctk.CTkFont(size=14),
+            height=40,
+            fg_color="green",
+            state="disabled",
+            command=self.export_model
+        )
+        self.export_button.pack(pady=5)
+
         # Progress
         self.progress_bar = ctk.CTkProgressBar(self.content, width=400)
         self.progress_bar.pack(pady=10)
@@ -1434,6 +1446,32 @@ class TrainingPage(BasePage):
 
         finally:
             self.after(0, lambda: self.train_button.configure(state="normal"))
+            self.after(0, lambda: self.export_button.configure(state="normal"))
+
+    def export_model(self):
+        """Export trained model to C header."""
+        if not self.classifier or not self.classifier.is_trained:
+            messagebox.showerror("Error", "No trained model to export!")
+            return
+
+        # Default path in ESP32 project
+        default_path = Path("EMG_Arm/src/core/model_weights.h").absolute()
+        
+        # Ask user for location, defaulting to the ESP32 project source
+        filename = tk.filedialog.asksaveasfilename(
+            title="Export Model Header",
+            initialdir=default_path.parent,
+            initialfile=default_path.name,
+            filetypes=[("C Header", "*.h")]
+        )
+        
+        if filename:
+            try:
+                path = self.classifier.export_to_header(filename)
+                self._log(f"\nExported model to: {path}")
+                messagebox.showinfo("Export Success", f"Model exported to:\n{path}\n\nRecompile ESP32 firmware to apply.")
+            except Exception as e:
+                messagebox.showerror("Export Error", f"Failed to export: {e}")
 
     def _log(self, text: str):
         """Add text to results."""
@@ -1861,115 +1899,158 @@ class PredictionPage(BasePage):
             self._update_connection_status("gray", "Disconnected")
             self.connect_button.configure(text="Connect")
 
-    def _prediction_thread(self):
-        """Background prediction thread."""
-        # For simulated mode, create new stream
-        if not self.using_real_hardware:
-            self.stream = GestureAwareEMGStream(num_channels=NUM_CHANNELS, sample_rate=SAMPLING_RATE_HZ)
+    def toggle_prediction(self):
+        """Start or stop prediction."""
+        if self.is_predicting:
+            self.stop_prediction()
+        else:
+            self.start_prediction()
 
-        # Stream is already started (either via handshake for real HW or will be started for simulated)
+    def start_prediction(self):
+        """Start live prediction."""
+        # Determine mode
+        self.using_real_hardware = (self.source_var.get() == "real")
+        
+        if self.using_real_hardware:
+            if not self.is_connected or not self.stream:
+                messagebox.showerror("Not Connected", "Please connect to ESP32 first.")
+                return
+
+            print("[DEBUG] Starting Edge Prediction (On-Device)...")
+            try:
+                # Send "start_predict" command to ESP32
+                if hasattr(self.stream, 'ser'):
+                   self.stream.ser.write(b'{"cmd": "start_predict"}\n')
+                   self.stream.running = True 
+                else:
+                   # Fallback
+                   self.stream.ser.write(b'{"cmd": "start_predict"}\n')
+                   self.stream.running = True
+                   
+            except Exception as e:
+                messagebox.showerror("Start Error", f"Failed to start: {e}")
+                return
+
+        else:
+            # Simulated - use PC-side inference
+            self.stream = GestureAwareEMGStream(num_channels=NUM_CHANNELS, sample_rate=SAMPLING_RATE_HZ)
+            self.stream.start()
+
+        # Load model for PC-side (Simulated) OR for display (optional)
+        # Even for Edge, we might want the label list.
+        if not self.using_real_hardware:
+            if not self.classifier:
+                model_path = EMGClassifier.get_default_model_path()
+                if model_path.exists():
+                    self.classifier = EMGClassifier.load(model_path)
+                    self.model_label.configure(text="Model: Loaded", text_color="green")
+                else:
+                    self.model_label.configure(text="Model: Not found (Simulating)", text_color="orange")
+            
+            # Reset smoother
+            self.smoother = PredictionSmoother(
+                label_names=self.classifier.label_names if self.classifier else ["rest", "open", "fist", "hook_em", "thumbs_up"],
+                probability_smoothing=0.7,
+                majority_vote_window=5,
+                debounce_count=3
+            )
+
+        self.is_predicting = True
+        self.start_button.configure(text="Stop Prediction", fg_color="red")
+        
+        # Start display loop
+        self.prediction_thread = threading.Thread(target=self.prediction_loop, daemon=True)
+        self.prediction_thread.start()
+        
+        self.update_prediction_ui()
+
+    def stop_prediction(self):
+        """Stop prediction."""
+        self.is_predicting = False
+        if self.stream:
+            self.stream.stop() # Sends "stop" usually
+            if not self.using_real_hardware:
+                self.stream = None
+                
+        self.start_button.configure(text="Start Prediction", fg_color=["#3B8ED0", "#1F6AA5"])
+        self.prediction_label.configure(text="---", text_color="gray")
+        self.confidence_label.configure(text="Confidence: ---%")
+        self.confidence_bar.set(0)
+
+    def prediction_loop(self):
+        """Loop for reading data and (optionally) running inference."""
+        import json
+        
         parser = EMGParser(num_channels=NUM_CHANNELS)
         windower = Windower(window_size_ms=WINDOW_SIZE_MS, sample_rate=SAMPLING_RATE_HZ, overlap=0.0)
 
-        # Simulated gesture cycling (only for simulated mode)
-        gesture_cycle = ["rest", "open", "fist", "hook_em", "thumbs_up"]
-        gesture_idx = 0
-        gesture_duration = 2.5
-        gesture_start = time.perf_counter()
-        current_gesture = gesture_cycle[0]
-
-        # Start simulated stream if needed
-        if not self.using_real_hardware:
-            try:
-                if hasattr(self.stream, 'set_gesture'):
-                    self.stream.set_gesture(current_gesture)
-                self.stream.start()
-            except Exception as e:
-                self.data_queue.put(('error', f"Failed to start simulated stream: {e}"))
-                return
-        else:
-            # Real hardware is already streaming
-            self.data_queue.put(('connection_status', ('green', 'Streaming')))
-
         while self.is_predicting:
-            # Change simulated gesture periodically (only for simulated mode)
-            if hasattr(self.stream, 'set_gesture'):
-                elapsed = time.perf_counter() - gesture_start
-                if elapsed > gesture_duration:
-                    gesture_idx = (gesture_idx + 1) % len(gesture_cycle)
-                    gesture_start = time.perf_counter()
-                    current_gesture = gesture_cycle[gesture_idx]
-                    self.stream.set_gesture(current_gesture)
-                    self.data_queue.put(('sim_gesture', current_gesture))
-
-            # Read and process
             try:
                 line = self.stream.readline()
+                if not line:
+                    continue
+
+                if self.using_real_hardware:
+                    # Edge Inference Mode: Expect JSON
+                    try:
+                        line = line.strip()
+                        if line.startswith('{'):
+                            data = json.loads(line)
+                            
+                            if "gesture" in data:
+                                # Update UI with Edge Prediction
+                                gesture = data["gesture"]
+                                conf = float(data.get("conf", 0.0))
+                                
+                                self.data_queue.put(('prediction', (gesture, conf)))
+                                
+                            elif "status" in data:
+                                print(f"[ESP32] {data}")
+                        else:
+                            pass
+                            
+                    except json.JSONDecodeError:
+                        pass
+                        
+                else:
+                    # PC Side Inference (Simulated)
+                    sample = parser.parse_line(line)
+                    if sample:
+                        window = windower.add_sample(sample)
+                        if window and self.classifier:
+                            # Run Inference Local
+                            raw_label, proba = self.classifier.predict(window.to_numpy())
+                            label, conf, _ = self.smoother.update(raw_label, proba)
+                            
+                            self.data_queue.put(('prediction', (label, conf)))
+
             except Exception as e:
-                # Only report error if we didn't intentionally stop
                 if self.is_predicting:
-                    self.data_queue.put(('error', f"Serial read error: {e}"))
+                    print(f"Prediction loop error: {e}")
                 break
 
-            if line:
-                sample = parser.parse_line(line)
-                if sample:
-                    window = windower.add_sample(sample)
-                    if window:
-                        # Get raw prediction
-                        window_data = window.to_numpy()
-                        raw_label, proba = self.classifier.predict(window_data)
-                        raw_confidence = max(proba) * 100
-
-                        # Apply smoothing
-                        smoothed_label, smoothed_conf, debug = self.smoother.update(raw_label, proba)
-                        smoothed_confidence = smoothed_conf * 100
-
-                        # Send both raw and smoothed to UI
-                        self.data_queue.put(('prediction', (
-                            smoothed_label,     # The stable output
-                            smoothed_confidence,
-                            raw_label,          # The raw (possibly twitchy) output
-                            raw_confidence,
-                        )))
-
-        # Safe cleanup - stream might already be stopped
-        try:
-            if self.stream:
-                self.stream.stop()
-        except Exception:
-            pass  # Ignore cleanup errors
-
     def update_prediction_ui(self):
-        """Update UI from prediction thread."""
+        """Update UI from queue."""
         try:
             while True:
                 msg_type, data = self.data_queue.get_nowait()
-
+                
                 if msg_type == 'prediction':
-                    smoothed_label, smoothed_conf, raw_label, raw_conf = data
-
-                    # Display smoothed (stable) prediction
-                    display_label = smoothed_label.upper().replace("_", " ")
-                    color = get_gesture_color(smoothed_label)
-
-                    self.prediction_label.configure(text=display_label, text_color=color)
-                    self.confidence_bar.set(smoothed_conf / 100)
-                    self.confidence_label.configure(text=f"Confidence: {smoothed_conf:.1f}%")
-
-                    # Show raw prediction for comparison (grayed out)
-                    raw_display = raw_label.upper().replace("_", " ")
-                    if raw_label != smoothed_label:
-                        # Raw differs from smoothed - show it was filtered
-                        self.raw_label.configure(
-                            text=f"Raw: {raw_display} ({raw_conf:.0f}%) → filtered",
-                            text_color="orange"
-                        )
-                    else:
-                        self.raw_label.configure(
-                            text=f"Raw: {raw_display} ({raw_conf:.0f}%)",
-                            text_color="gray"
-                        )
+                    label, conf = data
+                    
+                    # Update label
+                    self.prediction_label.configure(
+                        text=label.upper(),
+                        text_color=get_gesture_color(label)
+                    )
+                    
+                    # Update confidence
+                    self.confidence_label.configure(text=f"Confidence: {conf*100:.1f}%")
+                    self.confidence_bar.set(conf)
+                    
+                    # Clear raw label since we don't have raw vs smooth distinction in edge mode
+                    # (or we could expose it if we updated the C struct, but for now keep it simple)
+                    self.raw_label.configure(text="", text_color="gray")
 
                 elif msg_type == 'sim_gesture':
                     self.sim_label.configure(text=f"[Simulating: {data}]")
diff --git a/learning_data_collection.py b/learning_data_collection.py
index b08cbaa..96281e6 100644
--- a/learning_data_collection.py
+++ b/learning_data_collection.py
@@ -1757,6 +1757,132 @@ class EMGClassifier:
         print(f"[Classifier] File size: {filepath.stat().st_size / 1024:.1f} KB")
         return filepath
 
+    def export_to_header(self, filepath: Path) -> Path:
+        """
+        Export trained model to a C header file for ESP32 inference.
+
+        Args:
+            filepath: Output .h file path
+
+        Returns:
+            Path to the saved header file
+        """
+        if not self.is_trained:
+            raise ValueError("Cannot export untrained classifier!")
+
+        filepath = Path(filepath)
+        filepath.parent.mkdir(parents=True, exist_ok=True)
+
+        n_classes = len(self.label_names)
+        n_features = len(self.feature_names)
+
+        # Get LDA parameters
+        # coef_: (n_classes, n_features) - access as [class][feature]
+        # intercept_: (n_classes,)
+        coefs = self.lda.coef_
+        intercepts = self.lda.intercept_
+
+        # Add logic for binary classification (sklearn stores only 1 set of coefs)
+        # For >2 classes, it stores n_classes sets.
+        if n_classes == 2:
+            # Binary case: coef_ is (1, n_features), intercept_ is (1,)
+            # We need to expand this to 2 classes for the C inference engine to be generic.
+            # Class 1 decision = dot(w, x) + b
+            # Class 0 decision = - (dot(w, x) + b)  <-- Implicit in sklearn decision_function
+            # BUT: decision_function returns score. A generic 'argmax' approach usually expects
+            # one score per class. Multiclass LDA in sklearn does generic OVR/Multinomial.
+            # Let's check sklearn docs or behavior.
+            # Actually, LDA in sklearn for binary case is special.
+            # To make C code simple (always argmax), let's explicitly store 2 rows.
+            # Row 1 (Index 1 in sklearn): coef, intercept
+            # Row 0 (Index 0): -coef, -intercept ?
+            # Wait, LDA is generative. The decision boundary is linear.
+            # Let's assume Multiclass for now or handle binary specifically.
+            # For simplicity in C, we prefer (n_classes, n_features).
+            # If coefs.shape[0] != n_classes, we need to handle it.
+            if coefs.shape[0] == 1:
+                print("[Export] Binary classification detected. Expanding to 2 classes for C compatibility.")
+                # Class 1 (positive)
+                c1_coef = coefs[0]
+                c1_int = intercepts[0]
+                # Class 0 (negative) - Effectively -score for decision boundary at 0
+                # But strictly speaking LDA is comparison of log-posteriors.
+                # Sklearn's coef_ comes from (Sigma^-1)(mu1 - mu0).
+                # The score S = coef.X + intercept. If S > 0 pred class 1, else 0.
+                # To map this to ArgMax(Score0, Score1):
+                # We can set Score1 = S, Score0 = 0. OR Score1 = S/2, Score0 = -S/2.
+                # Let's use Score1 = S, Score0 = 0 (Bias term makes this trickier).
+                # Safest: Let's trust that for our 5-gesture demo, it's multiclass.
+                pass
+
+        # Generate C content
+        timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        
+        c_content = [
+            "/**",
+            f" * @file {filepath.name}",
+            " * @brief Trained LDA model weights exported from Python.",
+            f" * @date {timestamp}",
+            " */",
+            "",
+            "#ifndef MODEL_WEIGHTS_H",
+            "#define MODEL_WEIGHTS_H",
+            "",
+            "#include <stdint.h>",
+            "",
+            "/* Metadata */",
+            f"#define MODEL_NUM_CLASSES {n_classes}",
+            f"#define MODEL_NUM_FEATURES {n_features}",
+            "",
+            "/* Class Names */",
+            "static const char* MODEL_CLASS_NAMES[MODEL_NUM_CLASSES] = {",
+        ]
+
+        for name in self.label_names:
+            c_content.append(f'    "{name}",')
+        c_content.append("};")
+        c_content.append("")
+
+        c_content.append("/* Feature Extractor Parameters */")
+        c_content.append(f"#define FEAT_ZC_THRESH {self.feature_extractor.zc_threshold_percent}f")
+        c_content.append(f"#define FEAT_SSC_THRESH {self.feature_extractor.ssc_threshold_percent}f")
+        c_content.append("")
+
+        c_content.append("/* LDA Intercepts/Biases */")
+        c_content.append(f"static const float LDA_INTERCEPTS[MODEL_NUM_CLASSES] = {{")
+        line = "    "
+        for val in intercepts:
+             line += f"{val:.6f}f, "
+        c_content.append(line.rstrip(", "))
+        c_content.append("};")
+        c_content.append("")
+
+        c_content.append("/* LDA Coefficients (Weights) */")
+        c_content.append(f"static const float LDA_WEIGHTS[MODEL_NUM_CLASSES][MODEL_NUM_FEATURES] = {{")
+        
+        for i, row in enumerate(coefs):
+            c_content.append(f"    /* {self.label_names[i]} */")
+            c_content.append("    {")
+            line = "        "
+            for j, val in enumerate(row):
+                line += f"{val:.6f}f, "
+                if (j + 1) % 8 == 0:
+                    c_content.append(line)
+                    line = "        "
+            if line.strip():
+                c_content.append(line.rstrip(", "))
+            c_content.append("    },")
+            
+        c_content.append("};")
+        c_content.append("")
+        c_content.append("#endif /* MODEL_WEIGHTS_H */")
+
+        with open(filepath, 'w') as f:
+            f.write('\n'.join(c_content))
+
+        print(f"[Classifier] Model weights exported to: {filepath}")
+        return filepath
+
     @classmethod
     def load(cls, filepath: Path) -> 'EMGClassifier':
         """