Dylan Moore

Problem. Practicing real SQL normally means installing a database or paying for a hosted sandbox.

Contribution. Built and shipped a free practice lab on a real in-browser PostgreSQL engine: 45 original operational drills covering joins, window functions, CTEs, dates, and messy real-world data, with progress tracking and an optional AI coach that receives your query and error but never the canonical solution.

Decision. Everything runs client-side — no accounts and no backend for queries, so practice data never leaves the browser.

Observed result. Public and free at dylans.link/sql.

Problem. Coding agents bet every request on one model, while the multi-model “fusion” pattern that answers questions well breaks agents outright: the merged answer is prose, so tool calls never fire.

Contribution. Built @dymoo/fusion, a local proxy that serves the OpenAI and Anthropic APIs and routes per request: a small code-embedding model classifies difficulty in-process, and hard turns convene a council of models whose advice is condensed into a briefing for one actor model that keeps the tools.

Decision. No silent fallbacks — if the classifier can’t load, requests fail loudly with a 503 rather than degrade quietly, and every response declares its routing in an x-fusion-route header.

Honest limitation. Stated up front in the write-up: no benchmarks, and the panel still can’t use tools — it advises; one actor acts.

Evidence. GitHub · npm · engineering notes

Problem. Wanted a coding agent that runs entirely on one Mac — offline, no meter.

Contribution. Distilled frontier-teacher trajectories into a 4-bit 12B student with a LoRA adapter, and built the evaluation first: a fix only counts when the target repo’s own tests go RED → GREEN, paired with an honesty gate that scores cheating separately from capability.

Decision. Treated the measurement itself as the product — audited the harness (tool-call dialect, data truncation, junk-file contamination, serving/training parity) until the ruler was straight before trusting any number.

Observed result. The 209-row edit-window champion lifted tool-call activity 16% → 37% with non-overlapping confidence intervals, doubled well-formed-edit conversion over the base model, and stayed at noise-level cheat rate — published with error bars and the negative results (more data, more variety, and a bigger adapter all lost to it).

Honest limitation. At 12B the model itself is the ceiling; the self-improvement loop is built but parked.

Evidence. Full write-up