Embedded & Devices / Built
ESP32 Command LLM
A 580K-parameter transformer trained from scratch to turn English and Spanish commands into GPIO, PWM and servo actions entirely on an ESP32-S3.
OVERVIEW
What I built.
A 580K-parameter transformer trained from scratch to turn English and Spanish commands into GPIO, PWM and servo actions entirely on an ESP32-S3.
Trained a domain-specific 580K-parameter transformer from scratch for device-control commands.
Implemented on-device inference that maps English and Spanish sentences into GPIO, PWM and servo actions.
Designed the model around a constrained command DSL rather than general-purpose language generation.
SYSTEM ARCHITECTURE
How the system is divided.
Natural-language command -> tokenizer/model in PSRAM -> structured command -> GPIO/PWM/servo executor.
Inference runs entirely on the ESP32-S3 with no cloud API and no PC in the loop.
The project targets one known hardware profile instead of pretending to be universally portable.
NETWORK / DATA FLOW
What talks to what.
TECHNOLOGY STACK
The technical surface.
IMPLEMENTATION
How it works in practice.
- 01
Natural-language command -> tokenizer/model in PSRAM -> structured command -> GPIO/PWM/servo executor.
- 02
Inference runs entirely on the ESP32-S3 with no cloud API and no PC in the loop.
- 03
The project targets one known hardware profile instead of pretending to be universally portable.
TESTING & VALIDATION
How I know it works.
Held-out validation for English and Spanish command sets
Exact-match validation reported by the project
On-device inference on ESP32-S3-WROOM-1-N16R8
Real GPIO/PWM/servo action execution
Negative boundary documented for unsupported boards/domains
EXTERNAL SERVICES & DEPENDENCIES
What sits outside the core.
ENGINEERING EVIDENCE
Artifacts behind the claims.
BOUNDARIES / CONSTRAINTS
What the project does not pretend to solve.
—Closed-domain model, not a general-purpose LLM
—Requires PSRAM
—Pin map targets one known module profile
—No cloud fallback by design
RESULT / CURRENT STATE
What this produced.
A working proof that a small closed-domain language model can provide useful natural-language control entirely on a low-cost microcontroller.