AI Engineer — I build production AI systems end to end.
AI Agents · MCP Servers · RAG Systems · Conversational AI (chat & voice)
End-to-end AI systems, not demos. I architect for the constraints that matter in production — latency budgets, failure modes, and provider abstraction, so model and vendor decisions stay reversible instead of baked into the codebase.
- 🔌 MCP Servers & Integrations — Model Context Protocol servers connecting AI to real systems: APIs, databases, CRMs, internal tools. Proper auth, structured errors, safety controls on write operations.
- 🤖 AI Agents & Multi-Agent Systems — Agents that plan, call tools, and complete multi-step workflows, with execution tracing and per-run cost visibility.
- 🔍 RAG & Knowledge Systems — Ingestion, chunking, retrieval, and grounded generation with confidence scoring and human escalation rather than confident guesses.
- 💬 Conversational AI — Chat and voice across web, messaging platforms, and telephony, including real-time streaming speech pipelines.
- ⚙️ LLM Application Architecture — Provider abstraction, model routing by task complexity, evaluation, and cost management across hosted and self-hosted inference.
| Project | What it does | The decision behind it |
|---|---|---|
| MCP Forge | Generates complete, runnable MCP servers from an OpenAPI spec or a Python library | Generation is deterministic — audited templates, no LLM writes code. These servers get handed to agents that take real actions. |
| Sonari | AI voice agent builder — browser, embedded widget, shareable link, or a real phone number | Real-time voice is a latency problem first. Transcription starts mid-sentence; playback starts before the response is complete. |
| Lumio | Multi-tenant RAG chatbot platform — web, WhatsApp, Instagram, Slack, Email | Answering is a five-stage pipeline, not one model call. Low-confidence answers escalate to a human instead of guessing. |
| TaskForce | Multi-agent workflow platform with live execution tracing | Per-agent model routing — small models on mechanical steps, strong models only where output quality is the deliverable. The cost split is visible. |
| MarketMind | Market analysis across US and Indian markets | No language model touches the arithmetic. Scoring is deterministic Python; the LLM only explains the result. |
Each runs end to end with a test suite and a live deployment.
AI & Agents Python · MCP SDK · Anthropic · OpenAI · Ollama · Whisper · ElevenLabs
Backend & Data FastAPI · WebSockets · PostgreSQL · ChromaDB · SQLite · Redis · Celery
Infra Docker · Twilio · Jinja2 · Pytest