Hardware components | ||||||
 |
| × | 1 | |||
Software apps and online services | ||||||
| ||||||
Hand tools and fabrication machines | ||||||
| ||||||
ENGRAM is an offline edge intelligence stack built to turn an audited host-side reasoning surface into a real ESP32-C3 deployment artifact.
This project does not claim unrestricted MCU LLM inference. It focuses on a bounded, reproducible, and auditable board path with public evidence.
Why I Built ItI built ENGRAM because I care more about deployable and verifiable intelligence than flashy demos.
The goal is not only to improve behavior on a host machine, but to pack that behavior into something that can run on real edge hardware and still be audited and reproduced.
What Is PublicThe public repository includes:
- accepted host-side results
- replayable evaluation outputs
- firmware binaries
- board-proof artifacts
- raw board readback JSON
- comparator and audit results
The current public ESP32-C3 board proof reaches `249 / 642` on the published fixed-batch path, with `host_full_match = 642 / 642`.
The published artifact size is `1, 380, 771 bytes`, and the firmware size on board is `1, 784, 352 bytes`
How It WorksAt a high level, the pipeline is:
1. Start from a frozen parent comparison boundary.
2. Apply host-side repairs and acceptance gates.
3. Audit the upgraded host-side surface.
4. Pack the accepted result into a bounded ESP32-C3 artifact.
5. Flash the firmware, read back the board report, and verify alignment.
Honest BoundaryA key part of this project is keeping the claim boundary explicit.
The board-proof line demonstrates bounded crystallization correctness on a real ESP32-C3.
It is not presented as unrestricted open-input MCU language modeling, and keeping that boundary explicit is an important part of the project.
Current StatusThis Hackster project currently focuses on published repo artifacts, board metrics, and reproducible evidence rather than a polished video demo.
Collaboration / Custom DevelopmentI am open to custom development, technical consulting, and pilot collaborations around offline edge AI, ESP32-C3 deployment, auditable inference pipelines, and resource-constrained embedded intelligence.
If you are building edge devices, local decision systems, or bounded on-device AI products and want a practical implementation partner, feel free to contact me at [[email protected]].
Comments