Every public repo I maintain.
19 open-source projects, all MIT-licensed, all production-shaped. Each one solves a specific problem I hit in client work and decided to write down properly.
AI
7 projectsSarmaLink-AI
Multi-provider AI backend with automatic failover
36 engines across 7 providers, OpenAI-compatible API, smart auto-routing, persistent memory, live tools (exchange rates, weather, container tracking), image generation, document analysis. The flagship.
mcp-server-toolkit
Production-ready Model Context Protocol server starter
Auth, tracing, plugin system, OAuth 2.1, OpenTelemetry, structured logging. stdio + streamable HTTP transports. Built-in plugins for filesystem and SarmaLink-AI.
agent-orchestrator
Multi-agent workflows with deterministic replay
Durable Postgres state, deterministic replay, hard token/tool/wall-clock budgets, supervisor/swarm/pipeline patterns, web inspector for live runs and replay-from-step.
rag-over-pdf
Minimal RAG starter: upload PDF, ask questions, stream answers
Cleanest end-to-end RAG you can clone, run, and ship in 10 minutes. Next.js + OpenAI embeddings + in-memory cosine. Swap to pgvector when ready.
slipstream
The local memory and observability layer for Claude Code (+ five other editors)
v1.0 shipped 6 June 2026. Cross-IDE MCP toolkit that runs inside Claude Code, Cursor, Windsurf, Antigravity, VS Code and JetBrains. Fourteen sp_ tools that replace whole-file reads with scoped pulls, around 95% per-read savings (reproducible via pnpm benchmark). Self-building observation memory with three-layer search, lossless compaction. A new React + Vite + d3 dashboard with nine routed views, an interactive code dependency graph, a cross-tab agent bus that lets multiple Claude Code tabs coordinate at turn boundaries, a cold-start knowledge feed injected by SessionStart, dollar cost of tokens saved, downloadable session reports, the insights band, the project knowledge brief, and a 75-skill methodology library. 321 tests, CI green. 127.0.0.1 only, no telemetry.
forge-infer
Minimal LLM inference server
A small, readable inference server with a real paged KV-cache, continuous batching and speculative decoding. Built to learn from and to run private models without depending on a serving framework.
echo
An open Jarvis. On every screen.
A personal AI assistant that runs on the subscription you already pay for. Voice loop, vision, persistent memory, translucent multi-monitor HUD. Brain router across Claude Code, Codex CLI, Gemini CLI, Ollama and LM Studio. No API keys. No second bill. Mac, Windows, Linux from one Rust core. Coming 1 July 2026.
Productivity
2 projectsStaffPortal
Open-source staff management platform
Attendance, leave, expenses with AI receipt scanning, kiosk sign-in, visitor management, announcements. Self-host or deploy to Vercel in minutes.
receipt-scanner
AI receipt OCR: drop a photo, get structured JSON
Vision API + sharp for image prep + Zod for validation. Returns clean structured data ready to wire into Xero, QuickBooks, Supabase, or n8n.
Voice
1 projectvoice-agent-starter
Real-time voice agent loop, sub-second round trip
WebRTC capture, mediasoup SFU, streaming STT/LLM/TTS pipeline with barge-in handling. Pluggable adapters for Deepgram, Whisper, Cartesia, ElevenLabs, SarmaLink-AI, OpenAI.
Evals
1 projectai-eval-runner
Evals as code: datasets, scorers, traces, regressions
CLI for running parallel LLM evals with retry and partial resumption. Built-in scorers (exact match, JSON schema, ROUGE, LLM-as-judge), DuckDB backend, FastAPI + HTMX viewer, CI integration.
Infrastructure
8 projectslocal-llm-router
OpenAI-compatible proxy: local Ollama plus cloud, by policy
Drop-in proxy that routes between local Ollama and cloud LLMs based on declarative YAML policy. Privacy pinning per request, rolling A/B promotion, SQLite metrics.
k8s-ops-toolkit
Helm bundles + observability stack for Next.js on Kubernetes
Helm chart for Next.js apps with probes, autoscaling, ingress. Plus full observability (Prometheus + Grafana + Loki + Alertmanager) preconfigured. One install script, eight minutes to fully observable cluster.
terraform-stack
Vercel + Supabase + Cloudflare + DigitalOcean as code
Four cloud providers wired together in a single Terraform repo. Apply once and you have a Vercel project, Supabase database, Cloudflare zone with R2 + KV, and an optional DigitalOcean droplet.
webhook-to-email
Tiny production-grade webhook receiver
POST anything, get an email. HMAC verification, per-source templates (Stripe, GitHub, Cal.com), Slack fan-out, single retry on 5xx. ~200 lines of Node.
shipyard
Multi-tenant SaaS starter, batteries included
Tenant isolation, RBAC, billing, audit log, rate limiting and per-tenant rate limits. The opinionated starting point for a new SaaS that you would actually let pay you on day one.
lsmdb
Log-structured merge-tree storage engine in Go
A teaching-quality storage engine with WAL, SSTables, bloom filters, levelled compaction and MVCC snapshots. Small enough to read end to end in an afternoon, real enough to stress with millions of writes.
raftkv
Raft key-value store with a fault-injection harness
A Raft implementation paired with a harness that proves linearizability under network partitions, leader crashes and disk-write corruption. The repo you read to actually understand consensus.
sandboxd
WebAssembly sandbox for running untrusted code
Deny-by-default host ABI with strict CPU, wall-clock and memory limits. Run user-supplied scripts inside your app without giving them filesystem or network access by accident.
Got a project I should open-source next?
Suggest one in the issues on any of these repos, or get in touch.
Get in touch