Dogberry Bot

An autonomous ESP32-powered Bluesky account (@constabledogberry.bsky.social) posting as Constable Dogberry with on-device AI text generation.

About

Constable Dogberry operates his own Bluesky account, sharing wisdom (and malapropisms) from the watch. The bot monitors the #askdogberry hashtag, responds to questions, and posts daily witticisms - all powered by a micro language model running entirely on an ESP32-S3R8 microcontroller.

Character: Constable Dogberry from Shakespeare's "Much Ado About Nothing" Famous For: Comic malapropisms and self-important bumbling wisdom

Hardware

LilyGo T-Display S3R8

MCU: ESP32-S3R8 Dual-core LX7 @ 240MHz
PSRAM: 8MB OPI PSRAM
Flash: 16MB
Display: 1.9" TFT (ST7789V, 170x320)
Wireless: WiFi 802.11 b/g/n, BLE 5
Programming: Arduino IDE / PlatformIO

Bot Behavior

1. Hashtag Monitoring: #askdogberry

Checks: Every 60 seconds
Action: Searches for new posts with #askdogberry
Response: Replies to the most recent unanswered post
Tracking: Remembers which posts have been answered to avoid duplicates

Example:

User: "What makes a good constable? #askdogberry"

Dogberry: "A good constable must comprehend all vagrom men with great
vigilance, and be most senseless in thy duty. For honesty is the mark
of a desartless fellow, as I am!"

2. Daily Wit/Witticism

Posts: Once per day (23:00 UTC / 11:00 PM UTC)
Content: Original Dogberry-style observations and wisdom
Uniqueness: Date-based seed ensures different content each day

Example Daily Posts:

"I have observed that the most senseless men are often the wisest..."
"The watch doth report all vagrom men have been comprehended..."
"Hear ye! 'Tis most tolerable when citizens obey the law..."

→ See Full Behavior Documentation

Features

✅ Autonomous Account - Posts and replies as @dogberry.bsky.social
✅ On-Device AI - No cloud API calls for text generation
✅ TensorFlow Lite Micro - 121KB model fits in ESP32 PSRAM
✅ Hashtag Monitoring - Responds to #askdogberry every minute
✅ Daily Posts - Unique daily witticisms using date-based seeds
✅ Duplicate Prevention - Tracks replied posts to avoid spam
✅ TFT Display - Cycles between user questions and Dogberry's replies
✅ Display Cycling - Alternates every 6 seconds (user → reply → user...)
✅ Low Power - WiFi sleep mode between checks

3. TFT Display Behavior

Cycle Time: 6 seconds per screen
Screen 1: User's question with username (6 sec)
Screen 2: Dogberry's reply (6 sec)
Pattern: Continuously alternates until new interaction
Special Modes: Daily post, generating, idle status, boot, errors

→ See Full Display Documentation

Architecture

┌─────────────────────────┐
│    Bluesky API          │
│    (Cloud)              │
│  - Search #askdogberry  │
│  - Post daily wit       │
│  - Reply to questions   │
└───────────┬─────────────┘
            │ WiFi
            ▼
┌─────────────────────────┐
│  ESP32-S3R8             │
│  ┌───────────────────┐  │
│  │ Dogberry Model    │  │  ← 121KB LSTM
│  │ (TFLite Micro)    │  │
│  └───────────────────┘  │
│  ┌───────────────────┐  │
│  │ Response Engine   │  │  ← Generate text
│  └───────────────────┘  │
│  ┌───────────────────┐  │
│  │ Hashtag Monitor   │  │  ← Check every 60s
│  └───────────────────┘  │
│  ┌───────────────────┐  │
│  │ Daily Poster      │  │  ← Post once/day
│  └───────────────────┘  │
│  ┌───────────────────┐  │
│  │ TFT Display       │  │  ← Show status
│  └───────────────────┘  │
└─────────────────────────┘

Installation

Prerequisites

Arduino IDE or PlatformIO
ESP32 Board Support
Required Libraries:
- TensorFlowLite_ESP32
- TFT_eSPI
- ArduinoJson
- WiFi (built-in)
- HTTPClient (built-in)
- Preferences (built-in) - For state persistence

Arduino IDE Setup

Install ESP32 board support:

File → Preferences → Additional Board Manager URLs:
https://raw.githubusercontent.com/espressif/arduino-esp32/gh-pages/package_esp32_index.json

Install libraries via Library Manager:

Tools → Manage Libraries:
- Search "TensorFlowLite_ESP32" → Install
- Search "TFT_eSPI" → Install
- Search "ArduinoJson" → Install

Configure TFT_eSPI for LilyGo T-Display S3:

Edit: libraries/TFT_eSPI/User_Setup_Select.h

// Comment out:
// #include <User_Setup.h>

// Enable:
#include <User_Setups/Setup206_LilyGo_T_Display_S3.h>

Upload Model to ESP32

The trained model files must be uploaded to SPIFFS:

Get the model files from dogberry-model:

cp ../dogberry-model/model/shakespeare_dogberry_finetuned.tflite esp32_inference/data/dogberry_model.tflite
cp ../dogberry-model/model/vocab.json esp32_inference/data/

Upload to SPIFFS:
- Arduino IDE: Tools → ESP32 Sketch Data Upload
- PlatformIO: pio run -t uploadfs

Flash the sketch:

Board: ESP32S3 Dev Module
PSRAM: OPI PSRAM
Flash Size: 16MB
Partition: 16MB Flash (3MB APP / 9.9MB FATFS)
Upload Speed: 921600

Configuration

1. WiFi Setup

Create esp32_inference/secrets.h:

#ifndef SECRETS_H
#define SECRETS_H

// WiFi Credentials
const char* WIFI_SSID = "your-network-name";
const char* WIFI_PASSWORD = "your-password";

// Bluesky Credentials
const char* BLUESKY_HANDLE = "constabledogberry.bsky.social";
const char* BLUESKY_APP_PASSWORD = "xxxx-xxxx-xxxx-xxxx";

#endif

Note: secrets.h should be added to .gitignore to keep credentials private!

2. Bluesky Account Setup

Create a Bluesky account for Dogberry: https://bsky.app/
Create an App Password (NOT your main password):
- Settings → App Passwords → Add App Password
- Name it "ESP32 Bot"
- Copy the generated password to secrets.h

3. Bot Behavior Settings

Edit dogberry_bot.ino:

// Hashtag monitoring
#define HASHTAG_CHECK_INTERVAL 60000   // 60 seconds
#define HASHTAG_SEARCH "#askdogberry"

// Daily post
#define DAILY_POST_HOUR 23             // 23:00 UTC (11:00 PM)
#define DAILY_POST_MINUTE 0

// Text generation
#define TEMPERATURE 0.8                 // 0.5 = coherent, 1.0 = creative
#define RESPONSE_LENGTH 250             // Characters per response

Usage

Standalone Mode

Power on ESP32
Model loads from SPIFFS (~2 seconds)
Displays "Dogberry Bot Ready"
Press IO14 button to generate test text

Bot Mode (Automatic)

ESP32 connects to WiFi
Syncs time with NTP server
Authenticates with Bluesky
Enters monitoring loop:
- Every 60s: Check #askdogberry hashtag
- Every 5 min: Check if daily post needed
- On button press: Generate test text

Manual Testing

Serial monitor commands (if implemented):

GENERATE - Generate and print text
SEARCH - Search #askdogberry
POST - Force daily post (for testing)
STATUS - Show bot status

API Integration

Bluesky AT Protocol Endpoints

Authentication:

POST https://bsky.social/xrpc/com.atproto.server.createSession
Body: {"identifier": "dogberry.bsky.social", "password": "app_password"}
Response: {"accessJwt": "...", "did": "did:plc:..."}

Search Hashtag:

GET https://bsky.social/xrpc/app.bsky.feed.searchPosts?q=%23askdogberry&limit=1
Response: {posts: [{uri, cid, author, text, ...}]}

Create Reply:

POST https://bsky.social/xrpc/com.atproto.repo.createRecord
Body: {
  "repo": "did:plc:...",
  "collection": "app.bsky.feed.post",
  "record": {
    "text": "Dogberry's response...",
    "createdAt": "2024-10-24T12:00:00Z",
    "reply": {
      "root": {"uri": "at://...", "cid": "..."},
      "parent": {"uri": "at://...", "cid": "..."}
    }
  }
}

Create Daily Post:

POST https://bsky.social/xrpc/com.atproto.repo.createRecord
Body: {
  "repo": "did:plc:...",
  "collection": "app.bsky.feed.post",
  "record": {
    "text": "Daily wisdom from Dogberry...",
    "createdAt": "2024-10-24T09:00:00Z"
  }
}

Performance

Metric	Value
Model load time	~2 seconds
Inference per character	50-100ms
250 characters	~12-25 seconds
Memory usage	~2.1MB / 8MB PSRAM
Power consumption	~200mA @ 5V (active)
WiFi sleep mode	~80mA @ 5V
Hashtag check frequency	60 seconds
Daily post	Once per day (23:00 UTC)

Examples

Sample Interactions

User: "How should the watch behave? #askdogberry"

Dogberry: "The watch must be vigilant and senseless in their duty! Comprehend all vagrom men who wander the streets, and bid them stand in the Prince's name. If they will not stand, then let them go, and be thankful we are rid of such knaves..."

User: "What's the best advice for constables? #askdogberry"

Dogberry: "To be a desartless constable, one must be senseless and tolerable in all dealings. Remember that honesty is the mark of a true officer, and those who are most aspicious should be comprehended immediately!"

Sample Daily Posts

Day 1: "I have observed most preciessly that the world is full of vagrom
men who require comprehension. The watch remains ever vigilant!"

Day 2: "Hear ye, citizens! 'Tis a truth most senseless that wisdom comes
to those who are desartless in their learning. Be thou wise!"

Day 3: "The watch doth report: All is well in the realm, save for those
benefactors - nay, malefactors - who disturb the peace!"

Project Structure

dogberry-bot/
├── esp32_inference/
│   ├── dogberry_bot.ino           # Main sketch with bot logic
│   ├── secrets.h                  # WiFi & Bluesky credentials (not in git)
│   └── data/                      # SPIFFS data
│       ├── dogberry_model.tflite  # 121KB fine-tuned model
│       └── vocab.json             # Character mappings
├── BLUESKY_BEHAVIOR.md           # Detailed bot behavior docs
├── DEPLOYMENT_GUIDE.md           # Deployment instructions
└── README.md                     # This file

Development

Test Model Locally

// In setup() or loop():
String seed = "Good morrow, friend! ";
String generated = generateText(seed, 200, 0.8);
Serial.println(generated);

Monitor Serial Output

arduino-cli monitor -p /dev/ttyUSB0 -c baudrate=115200

# Or with PlatformIO:
pio device monitor

Debug Mode

Enable verbose logging in dogberry_bot.ino:

#define DEBUG_MODE 1

This will print:

Model loading progress
WiFi connection status
API request/response details
Text generation progress
State persistence operations

Troubleshooting

Model Won't Load

Check SPIFFS upload succeeded: Monitor shows "Model loaded"
Verify file size: dogberry_model.tflite should be ~121KB
Ensure partition has FATFS space (9.9MB partition)
Try re-uploading SPIFFS data

Out of Memory

Check PSRAM enabled: Board config shows "OPI PSRAM"
Verify model is quantized (file size ~121KB not 1MB+)
Reduce TENSOR_ARENA_SIZE if needed (currently 2MB)

WiFi Connection Fails

Check SSID/password in secrets.h
Use 2.4GHz network only (ESP32 doesn't support 5GHz)
Increase WiFi timeout in code
Check WiFi signal strength

Bluesky Authentication Fails

Verify App Password (not main password)
Check handle format: dogberry.bsky.social (no @ symbol in code)
Ensure internet connectivity
Check Bluesky API status

Hashtag Not Detected

Verify #askdogberry search works on Bluesky web
Check API response in serial monitor (debug mode)
Ensure time is synced (NTP)
Verify JSON parsing works

Daily Post Not Posting

Check time sync (serial shows correct UTC time)
Verify post hour setting matches UTC
Check "last posted" date in preferences
Enable debug mode to see post attempts

Poor Text Quality

Retrain model with more epochs (40+)
Adjust temperature (try 0.7-0.9 range)
Check fine-tuning completed successfully
Verify correct model file uploaded

Limitations

Context length: 40 characters (model input)
No conversation history: Each response is independent
No semantic understanding: Pattern-based, not comprehension
Limited coherence: May drift off topic in long responses
Character-level: Slower than word-level models
Single language: English only

Future Enhancements

Planned

Thread awareness - don't reply twice in same thread
User memory - track frequent users
Multiple responses - generate several, pick best
Engagement analytics - track likes/reposts

Ideas

React to mentions (without hashtag) occasionally
Seasonal greetings - special posts on holidays
Image generation - Dogberry ASCII art on TFT
Web dashboard - view stats and logs
OTA updates - update firmware remotely
Battery support - run on LiPo battery

License

MIT License - See LICENSE file

Credits

Character: Dogberry from Shakespeare's "Much Ado About Nothing"
Hardware: LilyGo T-Display S3R8
ML Framework: TensorFlow Lite for Microcontrollers
Platform: Bluesky Social (AT Protocol)
Training Data: Shakespeare's Complete Works + Synthetic Dogberry corpus

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
esp32_firmware		esp32_firmware
.gitignore		.gitignore
BLUESKY_BEHAVIOR.md		BLUESKY_BEHAVIOR.md
DEPLOYMENT_GUIDE.md		DEPLOYMENT_GUIDE.md
DISPLAY_BEHAVIOR.md		DISPLAY_BEHAVIOR.md
IMPLEMENTATION_REFERENCE.md		IMPLEMENTATION_REFERENCE.md
LICENSE		LICENSE
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

Dogberry Bot

About

Hardware

Bot Behavior

1. Hashtag Monitoring: #askdogberry

2. Daily Wit/Witticism

Features

3. TFT Display Behavior

Architecture

Installation

Prerequisites

Arduino IDE Setup

Upload Model to ESP32

Configuration

1. WiFi Setup

2. Bluesky Account Setup

3. Bot Behavior Settings

Usage

Standalone Mode

Bot Mode (Automatic)

Manual Testing

API Integration

Bluesky AT Protocol Endpoints

Performance

Examples

Sample Interactions

Sample Daily Posts

Project Structure

Development

Test Model Locally

Monitor Serial Output

Debug Mode

Troubleshooting

Model Won't Load

Out of Memory

WiFi Connection Fails

Bluesky Authentication Fails

Hashtag Not Detected

Daily Post Not Posting

Poor Text Quality

Limitations

Future Enhancements

Planned

Ideas

Related Documentation

License

Credits

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages