Agentic AI Governance & Observability
A control plane for autonomous AI coding agents.
Self-hosted governance and observability platform for Claude Code and Gemini CLI. It captures every prompt, tool call, model response and token cost into Langfuse via native hooks routed through durable Temporal workflows, enforces real-time policy that audits and blocks risky actions, and adds LLM-as-judge goal assessment, anomaly detection, extended-thinking capture and per-session data lineage, all without modifying the agents.
PythonGoTemporalLangfuseClickHouse+5
MSc Thesis · Metosin · grade 5/5 FinanceFile
Bookkeeping & VAT for Finnish sole traders.
Turns everyday income & expenses into the figures Finnish sole traders (toiminimi) actually file: OmaVero VAT returns, income statement, general ledger, balance sheet and a filing calendar. Multi-line, multi-VAT-rate transactions, receipt uploads, per-user isolation via Postgres RLS, EN/FI/SV interface.
Next.js 16React 19TypeScriptSupabasePostgreSQL+3
Personal · Live in production Echo Agent
A private, local-first personal AI agent.
A self-hosted 'Jarvis' you talk to over Telegram. Learns your habits, connects to email, calendar & tasks. Runs fully locally via Ollama for privacy, or on cloud LLMs (Groq, Gemini) for speed. Each install is fully isolated.
PythonLLMsAI AgentsOllamaGroq+2
Distributed GPT-2 Inference
GPT-2's 12 layers, split across worker nodes.
A distributed text-generation system that splits GPT-2's transformer layers across worker nodes using PyTorch RPC. A FastAPI master orchestrates nodes, serves a chat UI, and persists conversations in MongoDB, with Prometheus + Grafana observability.
PythonPyTorchPyTorch RPCFastAPIMongoDB+4
Distributed LLM Serving
Low-latency model serving over gRPC + Kafka.
Serves LLMs using gRPC for low-latency inter-service calls and Kafka to decouple producers/consumers for scalable asynchronous inference; service interfaces defined with Protocol Buffers.
PythongRPCApache KafkaProtocol BuffersMicroservices
ratekit
Rate limiting with exact retry-after answers.
Token bucket and sliding-window log with honest mathematics: the 429 middleware's Retry-After header agrees with the limiter's computed deficit rather than a made-up wait. Injectable clocks make every time-based test deterministic and instant, and per-key limiters isolate callers.
PythonFastAPIASGIpytestGitHub Actions
hookline
The receiving half of webhooks, done properly.
Verifies HMAC signatures with the timestamp bound into the signed message, so captured deliveries can be neither replayed past the freshness window nor re-stamped. Deduplicates at-least-once deliveries by event id, acknowledges fast, and drives every event to a terminal state with retries, exponential backoff and a queryable dead-letter list.
PythonFastAPISQLitepytestGitHub Actions
muxa
Multimodal orchestration with retries, fallback and an eval gate.
Routes mixed media (text, images, audio) to the right model task, runs an async worker pool with retry-once and a deterministic fallback so no asset is ever dropped, validates every response against a closed shape, and keeps a per-run route and cost ledger. A keyword-recall eval gate fails CI when quality drops. Zero runtime dependencies.
PythonasynciopytestGitHub Actions
Archive Agent
An LLM tagging agent with a real eval harness.
Tags media archives with a closed taxonomy through any OpenAI-compatible endpoint, including Ollama. Structured-output validation, retry-once with a deterministic rule fallback, local vector search, and an eval harness that scores precision/recall/F1 against a labelled sample and gates CI on minimum F1.
TypeScriptNode.jsvitestGitHub Actions
repo-sentry
Policy-as-code auditing for GitHub repos.
A CLI that audits GitHub repositories against a YAML governance policy: branch protection and required reviews, CODEOWNERS, Actions token permissions, secret scanning, Dependabot. Emits a JSON report and CI-gate exit codes so a policy regression fails the build. Offline pytest suite, GitHub Actions CI.
PythonGitHub REST APIPyYAMLpytest
wasm-boids
A C++ flocking sim, tested before it ships to the browser.
A C++17 boids simulation compiled to WebAssembly with Emscripten. A deterministic native test suite (bounds, speed and NaN invariants, cohesion and separation contracts) runs in GitHub Actions before every wasm build, and measured build-size optimisation cut the JS glue from 9,034 to 3,758 bytes.
C++17EmscriptenWebAssemblyGitHub Actions
Doctors on Hand
Find nearby doctors on live maps.
A Flutter app that helps patients find nearby doctors using live maps, geolocation and route polylines, backed by Firebase (Auth, Firestore, Storage).
FlutterDartFirebaseGoogle MapsGeolocation