Niaz T.

Three things I do best: voice agents that sound human, multi-agent orchestration that actually ships, and RAG systems that hold up in production.

Voice agents. People on the other end of the call don't realize they're talking to an AI. That takes a lot of small things working together: VAD endpointing so the agent starts responding before the caller finishes, barge-in so it stops when interrupted, filler words and breathing pauses between sentences, dynamic speaking rate that slows down for complex info, sentiment-driven tone shifts, voice cloning. Sub-300ms latency. Self-hosted models alongside cloud providers. SIP telephony for outbound calls, plus inbound with intent routing and human handoff.

Agent orchestration. Built a system with five specialist agents (voice, browser, RAG, MCP tools, scheduler) that coordinate through dependency DAGs and run in parallel. The agents work through real tools: browser automation via Chrome DevTools Protocol, MCP servers managing Google Calendar, Gmail, GitHub, PostgreSQL, Slack. Per-room credential isolation. Human-in-the-loop approval gates for sensitive operations like calls and emails. AI that takes real actions, not just answers questions.

Production RAG. I've shipped:

Legal RAG (FlawFinder.ai) indexing US police departments, zero-hallucination citations, paying subscribers
Scientific document RAG with hybrid search, multi-language support, figure extraction via vision models
GraphRAG with 40% better retrieval vs vector search alone
I also help teams whose RAG works in demos but struggles with real data. Usually it's chunking strategy, missing evaluation, or grounding gaps. I audit the pipeline and rebuild the parts that need it.

Beyond the AI layer, I build the infrastructure that keeps systems running when real users show up: Kubernetes, CI/CD, monitoring, evaluation harnesses. Most recently Infrastructure Lead at an AI startup where I built the full platform from scratch.

Day to day: Python (FastAPI, Django), LangChain/LangGraph, PostgreSQL, pgvector, Redis, Kubernetes, Docker, Terraform, AWS. Frontend when needed (React/TypeScript). 15+ years in software, last 3 on LLM applications.

Based in Madrid (GMT+2), flexible for EST overlap. Available now, async-first, same-day responses.