分享一个基于 ESP32-C3 的离线微型专家项目

[Kaihong Deng]

大家好，分享一个我们最近公开出来的项目。

我们在 ESP32-C3 上做了一个离线微型专家运行时，并把公开展示用的仓库整理出来了。先放仓库链接：

[https://github.com/Alpha-Guardian/Engram]

这个项目的目标，不是做一个“能在单片机上自由聊天的大模型”，而是想探索另一件事：

能不能把一段有用但边界明确的能力，压成一个可以离线运行、真实烧到板子里、出了问题还能回读审计的小运行时。

目前仓库里公开的主要内容包括：

- ESP32-C3 固件二进制
- 烧录脚本
- 板端报告回读脚本
- 板端运行结果和审计结果
- 方法边界和复现说明

这个公开仓库重点展示的是：
1. 离线运行
不依赖云端推理
2. 真实板端验证
不是只在 PC 上模拟，而是真正在 ESP32-C3 上跑过，并且可以从板子里把报告读回来
3. 可审计
仓库里放了回读脚本、原始结果和汇总结果，尽量把“到底跑了什么”说清楚
4. 边界明确
我们没有把它包装成“单片机上的通用大模型”
当前公开线更接近一个受限、离线、可验证的微型专家运行时

也提前说明一下边界，避免误解：

- 这不是开放输入的通用 MCU 推理
- 这不是“把 ChatGPT 缩小后直接塞进 ESP32”
- 而是把一段特定能力“结晶化”成一个可以落地部署的小运行时

如果大家点进仓库，建议按这个顺序看：

1. README
2. docs/REPRODUCE.md
3. results/board_proof/
4. docs/TRUST_AND_AUDIT.md

发这个帖主要是想和大家交流两个问题：
1. 在 ESP32 这类芯片上，除了通用模型压缩之外，这种受限能力运行时有没有现实价值？
2. 如果往产品方向走，大家觉得更适合哪类场景：
- 工业设备
- 智能家居
- 边缘控制
- 机器人本地行为模块
- 其他

如果大家感兴趣，后面我也可以继续补一些内容，比如：

- 更清楚的结构图
- 更小一点的 demo case
- 板端运行过程说明
- 从宿主能力到板端运行时的打包思路

[xelth.com]

Hi Kaihong, your post resonated a lot. We've been building something very close in spirit: TensorCollapse (https://github.com/xelth-com/TensorCollapse, MIT-licensed) — "collapsing" one narrow, well-bounded behavior into a tiny runtime that gets flashed to a real board (ESP32-C5 / ESP32-S31 in our case), with most of the effort going into verification: every artifact ships with a manifest (schema instance + sha256), and its replies/timings are diffed byte-for-byte against a host-side reference model in an instruction-set simulator before it ever touches hardware, then re-checked on the wire after flashing.

On your question 1 — from our experience, yes: on this class of chip a bounded, auditable capability runtime is worth more than compressing a general model. Determinism is what MCUs are good at, and it is exactly the thing open-ended inference gives up. Your board_proof / TRUST_AND_AUDIT direction is the part most projects skip, so it was great to see it front and center.

If anything in our repo is useful to you, please take it — the schema-to-RV32 emitter (pure Python, no compiler toolchain needed) and the ISS-based byte-exact checks are probably the most transferable pieces. Honest caveat from our side: we don't have real usage telemetry built in yet, so debugging still leans entirely on the oracle checks rather than field data — that's the next thing on our list, and I'd be curious how you plan to learn from deployed boards too.

On question 2: edge control and local robot behavior modules feel like the strongest fit to us — anywhere someone has to sign off on "what exactly does this device do".