Embedded Systems · Edge AI · Autonomous Agents
#include <stdio.h>
#include <stdint.h>
typedef struct {
const char* role;
const char* focus;
uint32_t coffee_consumed;
} Engineer;
int main() {
Engineer russell = {
.role = "Embedded Systems & AI Architect",
.focus = "Bridging bare-metal hardware with autonomous AI agents",
.coffee_consumed = 0xFFFFFFFF
};
printf("Building reliable tooling and bringing hardware to life.\n");
return 0;
}Building reliable systems where embedded intelligence, autonomous agents, and practical developer tooling meet.
| Area | Current focus |
|---|---|
| Embedded & edge systems | STM32, ARM Cortex-M/A, Renesas, RTOS, bare-metal C/C++, PCB design and hardware-in-the-loop testing |
| AI agents & retrieval | LangGraph, MCP, human-in-the-loop workflows, RAG, hybrid retrieval, VLM-assisted automation |
| Developer systems | Python, Rust, FastAPI, Docker, CI/CD, browser automation and observability tooling |
| Product interfaces | React, TypeScript and Tailwind CSS |
This section intentionally highlights only projects confirmed as original work or user-authored extensions. Repositories that GitHub identifies as forks, mirrors, or reference copies are excluded.
| Project | Original work focus | Core stack |
|---|---|---|
| RCclaw | User-authored multi-channel agent integration layer and messaging extensions, built around the OpenClaw ecosystem. | TypeScript · AI agents |
| manbo-robot-dog | STM32 quadruped robot with voice, sensor safety and DWT profiling compiled out of production firmware by default. | C · STM32 · Resource-aware firmware |
| OfficeMind-New | Local office automation with a single-model lite runtime, bounded token budget and fail-closed browser approval gates. | Python · FastAPI · Local AI |
| c-hacks-library | Low-level C utilities with documented numeric-boundary semantics, GCC/Clang regression checks and sanitizer validation. | C · Systems programming · Testing |
Extreme capability should come from disciplined software, not compulsory hardware escalation.
The maintained projects expose resource budgets, eliminate unnecessary resident work, and separate diagnostics from production paths. Claims are tied to source, tests, or board-side measurement protocols rather than decorative badges.
| Repository | Verifiable maintenance outcome |
|---|---|
| manbo-robot-dog | DWT CYCCNT instrumentation is compiled out in the default production build (PERF_PROBE_ENABLE=0), so performance diagnostics do not reserve runtime resources; a separate measurement build produces board-side evidence. |
| OfficeMind-New | A lite profile starts without LangGraph or RAG, uses one configurable local model endpoint, defaults to a bounded 1,024-token response budget, and routes high-risk browser actions to human review before model execution. |
| c-hacks-library | Numeric boundary defects were corrected and the repository gained a reproducible make demo test sanitize quality gate validated locally with GCC and Clang. |
| RCclaw | A zero-dependency check:lite-profile guard now verifies the existing skip-channel gateway, startup-memory, performance-budget and hotspot checks remain available; the runtime guide treats channels and heavy capabilities as explicit opt-ins. |
| OfficeMind-Hackathon | The router now has a lite profile that selects a configurable small text model and fails closed for vision/OCR rather than silently loading an optional large model; its selection rules are shell-tested. |
| Browser Operator | A zero-third-party diagnostic path now isolates --help/check/tabs from Click, aiohttp, Playwright and OpenAI. In 7 cold-process median runs, root help fell from 275.312 ms / 28,076 KB RSS to 87.031 ms / 14,496 KB RSS (−68.39% / −48.37%); raw samples, tests and method are versioned with the code. |
| manus-skills | Added a standard-library skill preflight, a resource-budget workflow and a report template, so capability modules are selected and installed individually instead of being loaded wholesale. |
| fate-symphony | A --lite renderer bounds preview buffers, caps output at 11,025 Hz and skips optional reverb; the output WAV contract is regression-tested. |
| fried-chicken-manager | Profit and sensitivity calculations now have a standard-library CLI that runs without Tkinter or Matplotlib; input validation and 25-scenario analysis are regression-tested. |
This group presents user-authored agent integrations, office-automation experiments, reusable AI skills and browser-session tooling. Each showcased project now defines an explicit capability budget rather than treating all models, browsers, channels or retrieval components as mandatory runtime work.
This group focuses on original work in microcontroller-based robotics and C-oriented systems tooling.
These repositories showcase original experiments in generative computing and practical analytics, each with a functional low-resource path that avoids external libraries or graphical processing when it is not required.
The visitor badge records profile-page requests. It is a page-view counter rather than a unique-visitor measure.
Portfolio: russellcooper.me
Ask me about: embedded architecture, AI agent design, PCB layouts and advanced web automation.