eval-driven-development

A guard-railed, closed-loop workflow for AI coding agents: live state bus + execution-level hard intercepts for Claude Code, Codex, and DeepSeek Harness (GitHub PR / GitLab MR). From step-level to requirement-level; eval-driven, spec-driven, human-in-the-loop.

Companion code for the talk "Managing Production Agents at Scale — from Chaos to Reliability". One Google ADK agent, three production failure modes: eval-driven development, resilience, and zero-trust on Vertex AI.

AI-augmented QA platform for spec-driven development and testing, RAG-grounded analysis, eval-driven development and contract validation across Python, Go, Rust and Solidity.

Eval-first plugin builder for Claude Code — the eval suite is the contract; green is the definition of done. Primitive decision records, generated eval suites with arming gates, sha256-frozen contracts, goal-loop builds in isolated worktrees, verified ships.

🇧🇷 marolaGPT + the fine-tune, the DSPy prompt compile, the local model image and the benchmark gate

Autonomous skill improvement loop for Claude Code plugins — inspired by Karpathy's autoresearch. Modify → evaluate → keep/discard → repeat until convergence. Zero-touch quality iteration at scale.

Production harness for a multi-agent BI system — eval-gated, guardrailed, cross-source-validated. LangGraph + hybrid RAG + FastAPI, live on AWS.

正解表なしで5つの性質からソートを採点するメタモルフィック・オラクル | Sort graded by metamorphic relations

非決定的なシャッフルをカイ二乗検定で採点する統計オラクル | Shuffle graded by Monte-Carlo statistical tests

LLM査読の検出力をラベル付き見本で採点するメタ評価ハーネス | Meta-eval: grading an LLM reviewer with labeled specimens

Multilingual GenAI evaluation service across 5 task types and 3 languages, with regression-trend dashboard

ランダム入力の集中砲火で「壊れない」を採点するファジング・オラクル | Robust parser graded by a fuzzing (implicit) oracle

答えが一意でない出力を仕様アサーションで採点するEDD実証 | Divisor finder graded by a spec-assertion oracle

Claude Codeエージェント定義の査読と、その検出力を採点するオラクル | Agent-spec review graded by a detection-power oracle

Multi-agent inspection pipeline for solar cell EL images: EfficientNet-B0 severity classifier + Qwen3-VL (Ollama) reasoning, served via FastAPI. 75.3% on a 20-criteria eval suite.

JS→TS移行を元JSとの差分テストで採点するEDD実証 | JS→TS migration graded by differential testing

Python 2→3移植をgolden比較で採点するEDD実証 | Py2→3 migration graded by golden-output comparison

Behavioral regression eval: measure whether your AI rules are still needed on current models | AIへのルールが現行モデルにまだ必要かを実測でふるい分ける行動回帰テスト＋統計オラクル

decode(encode(x))==x の往復プロパティで採点するEDD実証 | RLE codec graded by a round-trip property oracle

日本語→SQLをSQLite実行結果で採点する実行オラクルのEDD実証 | Text-to-SQL graded by an execution oracle