ESP32-S3 Sentinel
A $4 always-on ESP32-S3 sentinel that runs a local char-LSTM decision tier and only wakes a richer gateway when confidence drops. 11.6 tok/s H512 hardware-verified.
Overview
ESP32-S3 Sentinel is a two-tier, local-first edge-AI system.
- Tier 0 sentinel — a $4 ESP32-S3 Freenove WROOM N8R8 runs a 6.34M-parameter char-level LSTM every 5 s and emits a confidence score.
- Tier 1 gateway — a laptop, UNO Q, or Raspberry Pi runs a heavier model via Ollama, but only when the sentinel escalates.
If the sentinel is confident, it answers locally. If confidence drops below the threshold, it posts sensor context to the gateway over WiFi, receives a one-sentence verdict, and displays it. If the gateway is down, the sentinel keeps running locally. No cloud, no blocking, no lies about state.
What the numbers mean
| Number | What it is | Source |
|---|---|---|
| 11.6 tok/s | TinyStories-class H512 inference on real ESP32-S3 hardware | hardware verification run, 2026-07-03 |
| 6.34M params | char-LSTM, hidden=512, 3 layers | models/char_lstm_709ff8a9_meta.json |
| 4.8 MB | int8 weights + int4 recurrent weight file | models/char_lstm_709ff8a9.bin |
| 33% throughput lead | vs. the closest comparable ESP32 language-model demo | competitive_landscape_esp32_lm.md |
| ~$12 | sentinel-only BOM | README hardware table |
| ~$60 | full setup with one gateway host | README hardware table |
These are receipts, not estimates. The model was trained locally, quantized, and verified on a real ESP32-S3.
Quickstart (no hardware needed)
The repo ships a worked example that runs entirely on a laptop:
git clone https://github.com/RecursiveIntell/esp32-sentinel
cd esp32-sentinel
bash scripts/install.sh
bash examples/dht22_laptop_gateway/run.sh
It simulates a DHT22 reading, starts the gateway on port 8090, escalates as if the sentinel sent it, validates the response schema, and writes a JSONL receipt. The example exits 0 only if the full flow works.
Use cases
- Room safety sentinel — DHT22 reads temp/humidity every 5 s; the sentinel decides whether conditions are safe, comfortable, or out-of-range, and escalates when uncertain.
- Multi-sensor policy router — combine DHT22 + soil moisture + PIR; each sensor maps to a canonical prompt and gets its own verdict.
- Fail-open always-on display — SSD1306 OLED shows local verdict, gateway answer, or “local only” if the gateway is unreachable.
Architecture
- Firmware — Rust
no_stdfor Xtensa LX7, inesp32-s3/src/bin/sentinel_wifi.rs. Usesesp-hal,embassy,microflow, andesp-radiofor WiFi. - Gateway — Python HTTP service in
uno-q/server/sensor_ai_gateway.py. Endpoints:GET /health,POST /api/sensor_ai,POST /api/train_lstm. - Protocol — plain HTTP/JSON between tiers. No SDK lock-in; either tier can be replaced independently.
Design rules
- Local-first — the only network call is the optional gateway escalation.
- Fail-open — any error, missing dependency, or empty result exits cleanly without blocking the prompt.
- Receipt or it didn’t happen — every sentinel decision and gateway response is logged as structured JSONL.
- No autonomous writes — the model emits verdicts; a human or explicit config decides what to do with them.
Links
Have questions about ESP32-S3 Sentinel?
Try asking the AI assistant! Here are some ideas:
Related Projects
AiDENs
A closed-loop self-learning AI that audits its own knowledge graph, detects gaps, generates tasks, executes them via LLM, and captures results as provenance-attributed facts — all local-first, all autonomous.
Gloss
A local-first desktop knowledge application for grounded chat over personal documents, with notebook-local storage and source-linked answers.
Palisade
A native Linux firewall control surface for nftables with privilege separation, validation, snapshots, and rollback.