I build AI systems end to end — from retrieval and agent orchestration to evaluation, observability, tool interfaces, and developer workflows. My recent work focuses on production-minded LLM applications, multimodal agents, MCP infrastructure, and resource-efficient language models.
I like taking ambiguous problems, turning them into working systems, and then making the systems easier to inspect, evaluate, and improve. The interesting problems aren't the ones with clean benchmarks; they are the ones where latency, evaluation, and reliability decide whether a model actually ships or just demos well.
| Repository | Stars | Pull Request | Description |
|---|---|---|---|
| The-PR-Agent/pr-agent | ⭐ 13.2k | #3736 | Prepared and normalized outgoing comment bodies in AWS CodeCommit provider before calling publish_code_suggestions to prevent AWS API payload rejection. |
| huggingface/accelerate | ⭐ 9.9k | #4296 | Fixed a critical checkpoint resumption regression where skip_first_batches dropped drop_last, non_blocking, slice_fn, and mesh parameters during distributed DataLoader reconstitution. |
| Agenta-AI/agenta | ⭐ 4.8k | #5377 | Fixed logging keyword argument typo (exc_infp → exc_info) in authentication service that swallowed exception stack traces during auto-join failures. |
| taylorwilsdon/google_workspace_mcp | ⭐ 3.3k | #1190 | Allowed Google Chat send_message MCP tool to send plain text messages without triggering strict union schema validation errors on unused optional fields. |
| taylorwilsdon/google_workspace_mcp | ⭐ 3.3k | #1189 | Fixed Google Forms tool creation by mapping title to documentTitle and applying form descriptions through subsequent batchUpdate calls. |
| taylorwilsdon/google_workspace_mcp | ⭐ 3.3k | #1155 | Normalized input attendee dictionaries into valid Google Calendar API v3 structures, preventing 400 Bad Request exceptions on event creation. |
| pydantic/pydantic-settings | ⭐ 1.5k | #982 | Resolved platform-specific static type checking error on Windows (mypy --platform win32) by safely guarding Path.is_mount attribute resolution. |
| qdrant/qdrant-client | ⭐ 1.4k | #1493 | Enforced UpdateMode semantics in Qdrant Local client batch updates and prevented operations on deleted points from reviving ghost points. |
| Lightning-AI/litData | ⭐ 615 | #928 | Fixed streaming dataset checkpoint state synchronization by resetting wrapper sample and cycle counters upon reset_state_dict. |
Explore all 46 open-source pull requests across 21 repositories on my Portfolio.
EnterpriseRAG Cross-repository code intelligence for engineering teams. Index repositories, retrieve relevant symbols with hybrid search + graph expansion, rerank evidence, and return answers with validated source citations. Python FastAPI LangGraph Tree-sitter Qdrant BM25 Neo4j RAG
SmartGuard A multimodal incident-auditing agent that turns long CCTV recordings into searchable evidence. Frame-level VLM captioning, semantic search over footage, MCP tool-use, ffmpeg clip extraction. Multimodal AI VLM MCP FastAPI Pixeltable Opik ffmpeg React
LLM Observability & Eval Dashboard A self-hosted stack for seeing what LLM applications are doing after deployment — traces, evaluation scores, drift signals, and operational alerts in one place. OpenTelemetry LLM Evals DeepEval RAGAS ClickHouse Prometheus Next.js
Arc - Architecture Canvas A developer tool that starts with a system idea, finds structurally relevant open-source projects, turns their real file trees into diagrams, and lets engineers compose and critique an architecture before writing code. TypeScript React React Flow CLI GitHub Search LLM
Clipagent A video search and clip extraction tool. Upload videos, ask questions in natural language, and extract clips by text or image similarity. FastMCP Pixeltable FastAPI React Docker
Telecall An AI voice mobile-carrier assistant that replaces call-center agents. LangGraph + Groq + FastRTC + Twilio + Qdrant for plan inquiries, billing, and technical support. LangGraph Groq FastRTC Twilio Qdrant STT TTS
Advanced RAG From Scratch Building RAG 2.0 from scratch — from keyword search to MultiModal RAG. Python RAG Retrieval Multimodal
Forla An async-first framework for transparent agents, multi-agent systems, and deterministic workflows in Python. Visible agent loop, typed protocol, persistent memory, middleware. Python Agents Framework Async Multi-Agent
DigitalOcean MCP Server MCP server for DigitalOcean infrastructure — manage Droplets, databases, domains, regions, and account information through a structured tool interface. MCP TypeScript DigitalOcean API
Context Visualizer A local proxy for coding-agent model requests that visualizes context blocks, shows token counts and cost, and lets you edit blocks before sending. TypeScript Context Engineering Developer Tools
Fraud Detection System End-to-end ML fraud detection system with full MLOps practices — notebooks, modular source, tests, Docker, GitHub Actions. Python MLflow Docker CI/CD
Efficient InkubaLM for Edge and Mobile Inference eaarly stage BUILDING PHASE Independent research on redesigning InkubaLM's attention architecture (LLaMA → DeepSeek Multi-head Latent Attention) for fast on-device inference while preserving low-resource language knowledge through knowledge distillation.
LLM-Powered Android Keyboard Early-stage BUILDING PHASE experimentation on whether a small on-device language model can provide useful contextual next-word prediction within mobile compute, memory, latency, and battery constraints.
