/EdgeLog · Methods
Methods
Ways of working that carry over to any project, measured: AI agents, firmware, review loops, and the runs that failed.
- Tools
- EmbedEval harness-maker
15 essays · latest first
- 01How Many AI Code Reviews Is Enough?I spent 400 LLM calls on one codebase to learn this. One review pass finds 34% of the real defects, and the four most-repeated findings were all false.
- 02LLM PCB Routing: What I Learned From a Human's FixesI handed a 96% auto-routed PCB to a contract engineer. He redrew finished routes, not only the last 4%. I kept that difference as 16 rules for my agent.
- 03How Long Does Harness Engineering Have?Everything useful in a harness ends up absorbed into vanilla. What survives is what the vendor structurally will not or cannot build. Notes on the shift.
- 04What loop would keep my software improving with LLMs?A hypothesis I am still sorting out: whether I can wire my own project to LLMs and turn it into a loop that keeps improving itself.
- 05The CRA Begins: Embedded Software's Center of Gravity MovesSeptember 11, 2026: the EU Cyber Resilience Act's first duty applies. I see it less as compliance than as embedded engineering's center of gravity moving.
- 06AI-Written Firmware: What to Hand Over, Where to Stop It3-Tier Trust sorts PRs, then verification gates (spec, TDD, RAG, HIL, deterministic analyzer, more) take them to merge. Each catches its own defect class.
- 07AI-Written Firmware: Measuring the Gap With 233 TestsSonnet 4.6 scored 68.0% and Haiku 4.5 56.9% on 233 firmware cases. EmbedEval is my open benchmark: 5-layer eval, n=3, Wilson 95% CI, Zephyr, STM32 HAL.
- 08AI-Written Firmware: What Breaks Outside the FunctionAI-written firmware breaks where a single function can't show it. I found 3 patterns visible in the diff and 3 that only surface after 49 days.
- 09AI-Written Firmware: What I Didn't Tell ItA junior wrote a logging daemon with an AI tool. It passed my review, ran fine, and later wore out an eMMC. The risk was in what the prompt left out.
- 10AI Finally Needs Hardware EngineersIn 2026 inference overtook training. The edge is where it grows fastest, and the work left there, I think, is the work AI cannot do.
- 11Offline Datasheet RAG: The Pipeline I Argued AgainstI argued against RAG for hardware docs, then an offline job made me build one. What survived: SVD parsing, structural chunks, hybrid search, multi-query.
- 12AI Agents in Embedded: Opening the Full PipelineI open one AI-agent-built i.MX8M Plus sensor driver project, folder by folder: specs, property tests, review agents, and feedback loops.
- 13AI Agents in Embedded: Learning Context EngineeringAI agents wrote good Yocto recipes but silently corrupted sensor data and misconfigured SPI registers. What fixed it was the context I gave them.
- 14AI Agents in Embedded: The Spec-Driven TDD We Put OffEmbedded teams skipped TDD for its cost. With AI agents I watched that math change. Notes on testing a Yocto IMU driver from a spec, with JSON feedback.
- 15Hardware Docs for LLMs: Why I Went Back to grepRAG cut the cross-references that make hardware docs useful. After wrestling with a pipeline, I went back to handing the LLM my SVD and PDF as file tools.