EdgeLog개인How Many AI Code Reviews Is Enough?
400 LLM calls on one codebase. A single review pass finds 34% of the real defects, and the four most-repeated findings were all false positives.
EdgeLog
개인최근 글
LLM PCB Routing: What the Contract Engineer Taught It
Seven of the sixteen PCB routing techniques in my agent's registry, most of them learned from a contract engineer, are ones it cannot execute.
Per-Project AI Agent Harness: Side vs Production
I run side and production projects at once, so the workflow has to differ. Most harnesses are fixed, so I built one that generates a harness per project.
SpotOn: How I Run an AI Team of One
Most of the first draft is written by an AI team. But the parts I have to know, and the ones I decide, I understand end to end. Read wide, write alone.
SpotOn: The Two Seats a Harness Can't Fill
I re-sorted the twenty-eight disciplines a hardware product needs. Twenty are files, so the harness owns them. Six stop at a doorway. Two stay human.
SpotOn: A Roadmap One Person Can Finish
A solo founder's roadmap isn't a build schedule. It's a map of kill-gates: cheap checkpoints before the expensive stages, and permission to quit early.
SpotOn: The Co-Founder That Builds Too Fast
The harness made building so cheap that I built almost the whole product before validating anyone will pay. The dangerous side of a good co-founder.
SpotOn: What I'm Building, and Why Alone
A tennis player and 18-year embedded engineer builds an impact sensor solo, with an LLM as the team. The start of a productization diary, kept either way.
How Long Does Harness Engineering Have?
Everything useful in a harness eventually gets absorbed into vanilla. What survives is only what the vendor structurally will not or cannot build.
What loop do I need to keep improving my software with LLMs?
Working hypothesis on whether I can wire my own project into an LLM-driven self-improvement loop.
The CRA Begins: Embedded Software's Center of Gravity Shifts
September 11, 2026: the EU Cyber Resilience Act's first obligation hits. Not compliance work. A center-of-gravity shift in embedded software engineering.
AI-Written Firmware: How to Stop AI from Breaking Firmware
3-Tier Trust classifies PRs. Verification gates (spec, TDD, RAG, HIL, deterministic analyzer, more) take them to merge. Each catches a different defect class.
AI-Written Firmware: 233 Tests, Sonnet 68% vs Haiku 57%
Sonnet 4.6 hits 68.0% and Haiku 4.5 56.9% on 233 embedded firmware cases. EmbedEval is open source: 5-layer eval, n=3, Wilson 95% CI, Zephyr and STM32 HAL.
AI-Written Firmware: Where It Breaks
AI-generated firmware fails in predictable nonlocal patterns. 3 visible in diff, 3 that detonate after 49 days.
AI-Written Firmware: How Far Can You Trust It?
I asked an LLM to write a logging daemon. It compiled, ran fine, and was silently killing the eMMC. What you leave out of the prompt is where embedded AI code breaks.
AI Finally Needs Hardware Engineers
Inference overtook training in 2026. The edge is growing fastest, and the work that remains is the work AI cannot do.
Offline Datasheet RAG: What Survived the Pipeline
Default RAG breaks on hardware docs. Here is the pipeline that works offline: SVD parsing, structural chunks, hybrid search, multi-query decomposition.
AI Agents in Embedded: The Full Pipeline
Complete project anatomy for AI-agent-driven embedded development. Specs, property tests, review agents, and feedback loops on i.MX8M Plus.
AI Agents in Embedded: Context Engineering
AI agents write good Yocto recipes but silently corrupt sensor data and misconfigure SPI registers. The fix is context engineering.
AI Agents in Embedded: Spec-Driven TDD
Embedded teams skipped TDD because the cost was too high. AI agents change the economics. Spec-driven testing for an IMU driver on Yocto.
Hardware Docs for LLMs: grep vs RAG
RAG destroys the cross-references that make hardware docs useful. Skip the pipeline: give the LLM your SVD and PDF as file tools.
Why I Started EdgeLog
What happens when you push LLMs, zero-trust security, and modern frameworks onto real embedded hardware? A research log from the edge.