I build things I can actually check, not things I merely believe work: a chess engine whose move generator is verified against perft counts, a Tic-Tac-Toe AI proven unbeatable by an exhaustive game-tree test, a 15-puzzle generator that only ever emits solvable boards because it accounts for the parity invariant that makes half of all tile arrangements unreachable.
That same test-first, verify-after habit carries into a newer, larger focus: tooling that makes LLM and agent systems cheaper to run, easier to debug, and safer to trust — cost calculators, RAG evaluators, prompt-regression checks, and agent-transcript forensics, built the same way as the games: with a documented way to check the claim, not just a README that asserts it.
| Pillar | Count | What it is |
|---|---|---|
| Flagship projects | 5 | The projects I'd point you to first |
| LLM & agent-ops toolkit | 24 | Cost, RAG-quality, prompt-testing, and agent-debugging tools |
| MCP servers | 4 | Small, focused servers for ops/automation and OSS triage |
| 58 Flutter games | 58 | Classic games, each with a real algorithm underneath, not a stub |
- Test-first development — unit tests, exhaustive/property checks, and reference-verified algorithms (perft, parity invariants) where a simple pass/fail isn't enough
- LLM/agent cost governance: pricing comparison, cache-yield estimation, spend anomaly detection
- RAG pipeline evaluation: retrieval-method benchmarking, groundedness checking, index-staleness monitoring
- Cross-platform app development: Flutter/Dart, Android/Kotlin
- Reproducible builds and releases (Gradle/Android CI, versioned GitHub releases)
24 small, single-purpose Python tools. Each does one job for LLM/agent systems in production — no dashboards, no SaaS, just a CLI you point at your own logs.
Cost & pricing
| Repo | What it does |
|---|---|
| batch-economics | Real-time vs. batch-API tradeoff calculator — accounts for the latency cost batching hides |
| cache-yield | Estimates how much prompt caching would've saved, from a historical request log |
| finetune-vs-prompt | Cost breakeven calculator: fine-tuning vs. long-context/RAG prompting |
| model-arbitrage | Cross-provider (Anthropic/OpenAI/Google) pricing comparator for a fixed workload shape |
| pareto-router | Finds the cost/quality Pareto frontier in a log of real request outcomes |
| rag-costmap | Projects RAG indexing, storage, and query costs from 10K to 10M docs |
| spend-sentinel | Statistical anomaly detector for LLM billing/usage logs |
| tokendiff | CI check that fails the build when a prompt template's token cost regresses |
| tokenledger | Token cost-attribution ledger for LLM API usage |
RAG & retrieval quality
| Repo | What it does |
|---|---|
| chunk-bench | Benchmarks document-chunking strategies by their effect on retrieval recall@k |
| context-packer | Picks the max-relevance subset of retrieved chunks that fits a token budget |
| groundcheck | Sentence-level groundedness/hallucination checker against retrieved context |
| index-decay | Staleness/drift monitor for RAG vector indexes via content-hashed manifests |
| query-rewriter-bench | Benchmark harness for RAG query-rewriting strategies |
| retrieval-arena | Compares sparse (BM25), dense, and hybrid retrieval on the same corpus and eval set |
Prompt & guardrail testing
| Repo | What it does |
|---|---|
| guardrail-bench | Adversarial red-team harness scoring a system prompt's guardrails via an LLM judge |
| prompt-regress | Behavioral regression testing for prompt templates — catches semantic drift and format breaks |
| rubric | LLM-as-judge eval harness — rubric criteria in, structured pass/fail report out |
| tool-contract | Static schema-drift detector for MCP/Claude tool JSON Schema definitions |
Agent observability & debugging
| Repo | What it does |
|---|---|
| agent-graph | Reconstructs a multi-agent call tree from a transcript and shows where tokens/cost went |
| session-replay | Deterministic offline replay/debugger for agent transcripts — diffs two runs to find where they diverged |
| waypoint | Durable, checkpointed workflow engine for chained agent steps, with human-approval gates |
| workflow-profiler | Per-step flame-graph profiler for agent workflow time and token cost |
Other
| Repo | What it does |
|---|---|
| flowforge | Config-driven, pluggable streaming ETL pipeline toolkit for Python |
| Repo | What it does |
|---|---|
| triagekit-mcp | TriageKit — a suite of small, focused MCP servers for automation engineers |
| pipeline-doctor | Diagnoses Airflow/dbt run failures, traces lineage impact, catches upstream-caused false alarms |
| k8s-triage | Diagnoses unhealthy Kubernetes workloads and finds right-sizing candidates from live cluster state |
| oss-scout | Finds genuinely-contributable open source issues — the unclaimed kind, not the ones under six duplicate PRs |
Each one is a full implementation with a real algorithm behind it, not a tutorial stub — a representative sample:
| Game | What makes it real |
|---|---|
| chess | Move generator verified with perft against chessprogramming.org reference values |
| tictactoe | Minimax AI verified unbeatable by an exhaustive never-loses test |
| gomoku | Alpha-beta minimax with an open-run-weighted threat-scoring heuristic |
| lightcycles | AI opponent flood-fills reachable space each tick — a real Tron-bot technique |
| fifteenpuzzle | Generated boards obey the parity invariant, verified against the historic unsolvable 14-15 swap |
| freecell | The actual ruleset — single-card moves aren't a simplification, that's how Freecell works |
| pixelchase | Pac-Man-style ghosts recompute a BFS shortest path every tick, then flee once you're powered up |
| tetris | 7-bag randomizer; tetromino rotations derived by rotating a matrix, not hand-copied tables |
Open an issue on any repo to start a conversation — label it "collab". If you prefer private contact, tell me which contact method to include.