Projects I build in the open

Self-hosted Go LLM gateway with a correctness-gated cache and per-run cost attribution

Self-hosted LLM gateway in Go that speaks OpenAI's Chat Completions and Embeddings wire format, plus Anthropic's Messages API, in front of OpenAI, Anthropic, Gemini, AWS Bedrock and any OpenAI-compatible server. One binary carries a three-layer response cache whose near-duplicate hits must pass non-configurable correctness gates, per-tenant virtual keys with Decimal budgets and Redis-backed rate limits, regex guardrails, and OpenTelemetry spans that attribute cost to an agent_run_id. A separate Python CLI, evals, scores agent outputs with Wilson intervals and an opt-in two-judge panel. 23 gateway releases through v0.20.0 (2026-10-15); solo-maintained.

Approve one exact MCP tool-definition digest; the gateway lists only tools that match it

Go control plane against MCP tool poisoning by rug pull. It crawls each upstream MCP server, versions every tool definition by the SHA-256 of its RFC 8785 canonical JSON (`_meta` included), and lets a named person approve one exact digest. It runs as agentgateway v1.6.0's ExtMCP processor, so a tool is listed only while its live digest still equals the approved one, and a call is denied whenever the evidence is missing or disagrees. Every decision lands in a hash-chained SQLite audit log, and since v0.4 human decisions are signed with OpenSSH keys. Single-tenant, CLI-only, pinned to one gateway version, one author.

Fail-closed query-cost guardrails for AI agents on BigQuery, Snowflake and Databricks

A single-process stdio MCP server that prices a SQL statement before an AI agent runs it: BigQuery dry run (PRECISE), Snowflake EXPLAIN (UPPER_BOUND), Databricks EXPLAIN COST (HEURISTIC). It refuses to execute when the estimate exceeds the caller's cap or when no comparable figure exists. Four tools, a schema-enforced accuracy tier on every estimate, read statements only by default, and results capped at 1000 rows and 8 MiB and labelled untrusted data. Ten PyPI releases in 37 days, the last three remediating a security audit, with SLSA provenance and a hash-pinned build. Solo-maintained; one GitHub star.

Recreation of Stripe's four-tier rate limiter and load shedder, as a Go sidecar and core

Open-source recreation of the four-tier rate limiter and load shedder Stripe described in a 2017 engineering post: a per-key token bucket, a per-key concurrent-request cap, a fleet-wide shedder that reserves capacity for critical traffic, and a sidecar-local in-flight shedder. Thin Go and Python HTTP clients call a per-host Go sidecar, which forwards checks over gRPC to a central Go core that keeps shared state in Redis through atomic Lua scripts. Every tier has a shadow mode. One author, 22 releases through v2.13.1 (2026-09-02).

Read-only MCP server for querying LLM traces across eight OpenTelemetry backends

Fork of traceloop/opentelemetry-mcp-server grown into an independently released MCP server for debugging LLM traces: 8 OpenTelemetry backends (Jaeger, Tempo and Traceloop inherited; Datadog, Sentry, AWS X-Ray, New Relic and Honeycomb added) behind one BaseBackend interface, 19 read-only tools (11 inherited) including deterministic root-cause triage and best-effort cross-backend correlation, gen_ai.* attribute parsing with litellm-backed cost attribution, a TTL cache with request coalescing, and SSRF, redirect and response-size guards. At v0.13.0 it is on PyPI, the MCP registry and GHCR; 303 of its 322 commits post-date the fork.

Production intelligence layer for 5,000+ feature flags.

Self-hosted feature flag intelligence with blast-radius gating, circuit-breaker auto-rollback, causal incident correlation ("What Changed?"), and Knight Capital–style tombstoning. Go + Python + TypeScript polyglot.

Agentic-intelligence framework for Claude Code.

Command-driven multi-agent orchestration for Claude Code — 221 commands, 164 subagents (152 adapted from VoltAgent, MIT), 354 skills, and cost-aware model routing. Ships as npx mindforge-cc.

Self-improving AI agent infrastructure.

A Karpathy-style propose/eval/score/commit-or-revert loop over a git-versioned agent spec, with an LLM-judge eval harness. Framework-agnostic.

Intelligence layer for autonomous AI agents.

SQLite/vector indexing, multi-agent orchestration, and resilience — shipped as core/CLI/MCP npm packages plus a 3D spatial dashboard.

Offline code intelligence engine — Tree-sitter indexing, hybrid BM25+vector+grep search

Zero-infra code intelligence — Tree-sitter AST indexing + up to 7-leg RRF retrieval (BM25 + vector + grep + SPLADE + multi-granularity; the graph-BFS leg is opt-in and measured harmful) + knowledge graph (Louvain community detection, Pyvis). MCP server. 4,435 unit test functions on main at v3.4.3. Works offline with no API key.

AI-native distributed code-intelligence platform.

Models code as a Neo4j knowledge graph + Chroma embeddings over 19 polyglot microservices, powering RAG over massive codebases with full observability.

AI-native bash interpreter, in TypeScript.

A sandboxed bash interpreter for AI agents — Tree-sitter WASM parser, in-process WASM runtimes (CPython, QuickJS, SQLite3), exposed over MCP.

Claude Code mod: one press rewrites a vague prompt from your CLAUDE.md and skills

A Claude Code mod, an in-process hooks module rather than a skill, that draws an Enhance / Send / Undo band above the prompt. One press sends the draft, the instruction files the session already loaded and the skill listing to a three-rung model chain with per-rung fallback, highlights the inserted words, and prints the call's cost on the band. Nine post-call guards refuse rewrites that add commands, invent paths or flip intent, and optional session and day spend caps sit on a per-writer ledger. MIT, 178 commits at v0.6.0, installed from GitHub as a plugin marketplace; not on npm.

Portfolio with four switchable views and a read-only MCP server over one MDX source

A personal engineering portfolio with four switchable experiences over one Velite content source: a static Classic site, an AI concierge chat grounded on the résumé corpus through AWS Bedrock with a model fallback chain and a deployment-tagged FAQ cache, a WebGL build graph, and a keyboard-native developer terminal. Agents read the same content through a read-only MCP server, llms.txt and resume.json. A fingerprinted citation gate, a hash-chained claims ledger and a first-load bundle budget check the docs and numbers in CI; none is a required merge check.

Notes, topics, or code → published technical articles.

STORM two-stage pipeline (outline → draft → polish) with BM25 in-memory RAG, streaming web UI, and one-command publishing to Dev.to (API), plus Medium and Hashnode publishers that script your own logged-in Chrome (not live-verified); Substack and LinkedIn stay manual.

Universal AI command manager.

Indexes 350+ slash commands, skills, agents, and plugins across Claude Code, Cursor, Copilot, Windsurf, and Aider into one searchable vault.

AI-curated daily tech digest: four sources, one batched Bedrock call, per-run cost ledger

A cron-driven Astro static site that pulls Hacker News, arXiv, GitHub and Dev.to through free keyless APIs, scores all twenty daily items in one forced-tool-use AWS Bedrock call (Sonnet 5 leading a four-model fallback chain), and validates the output against one Zod schema shared by the pipeline and the Astro content collection. GitHub Actions commits the result back to `main` each night. Every run's token usage and dollar cost goes into a committed JSONL ledger (52 runs, $3.10 total at 2026-10-20) with a 5x trailing-average anomaly flag that files a GitHub Issue.

Local-first CLI that turns one Markdown file into a presentable slide deck

One Markdown file in, a slide deck out: served on 127.0.0.1 with live reload, presented in the browser, or exported to PDF, per-slide PNGs or a static index.html, all from a single generateHtml() pipeline. Slides split at marked's hr tokens, KaTeX fonts ship as base64, the Google Fonts @import is stripped from every Mermaid SVG, and PDF export drives a Chrome-family browser the user already has through puppeteer-core, so there is no bundled browser, no CDN and no telemetry. 13 ADRs and 563 Vitest tests back it. The repo is tagged v1.10.0; npm still serves 1.7.0.