Hisan N.

Most voice agents and RAG bots work in the demo. Then real callers interrupt, real documents get messy, and the agent starts dropping leads or answering with confidence and no source. I build the version that holds up in production. Top Rated Plus · 100% Job Success · 50+ projects · 50+ five-star reviews I build production AI systems that keep working after launch: voice agents that handle live calls, RAG systems that answer from your own data with sources, and the backend automation that connects both to the tools you already run. Shipped across healthcare and dental, real estate, localization and coaching. Not prototypes. Every build includes structured data capture, error handling, logging, evals and a clean handoff to a human. RECENT WORK 📞 AI receptionist that answers every call, books the appointment and updates the CRM (real estate) Qualifies buyers on the call, books against live availability, writes the outcome to the CRM and warm-transfers to a human when qualification fails. Retell, ElevenLabs, Twilio SIP, FastAPI on Railway. 🎙 Voice assistant callers can interrupt, with answers pulled from your live data Sub-second turn-taking on the OpenAI Realtime API, streaming speech in and out, function calling for live lookups and barge-in handling. Client review: "Terrific collaborator. Smart, agile, curious, and adaptive." 🧠 AI coach your employees can talk to inside Microsoft Teams Built for a workplace-coaching company: voice conversations, document RAG and memory that carries across sessions, delivered as a Teams app. FastAPI on Azure with authentication, PostgreSQL, monitoring and locked-down infrastructure. (Amiko Coach) 🦷 Sourced answers from years of patient consultation notes, in seconds (dental) Staff ask in plain language across 1,400 consultation transcripts and get grounded answers with the source evidence shown. Hierarchical retrieval on Qdrant, hybrid semantic + BM25 search, metadata filtering, FastAPI ingestion and search backend. Consulting and architecture role. Client review: "Super happy with Hisan's work and will definitely re-engage him." 🔎 Prospect research and lead scoring on autopilot, straight into your CRM Agents find, enrich and score prospects, push qualified ones into the CRM and trigger follow-up. Duplicate prevention and retries on failed API calls so nothing drops silently. 🤝 Automatic user-to-user matching that runs in the background of a community platform Live for 1,000+ users: vector-similarity matching, LLM scoring of match quality, scheduled background searches, location filtering and automated onboarding. 🌐 Document translation on autopilot: upload Word, PowerPoint, Excel or PDF, get the finished file back Built for a professional translation company in Japan. Document processing, translation-memory and terminology retrieval, a three-pass LLM workflow, collaborative review, reconstruction of translated Word, PowerPoint, Excel and PDF files, and locked-down S3 storage. Four milestones delivered, ongoing. 🔒 Private ChatGPT-style assistant on your own cloud, nothing leaves your servers Self-hosted vLLM and RAG on RunPod and AWS for teams that can't send data to a third-party API. Includes an n8n stack migrated to self-hosted RunPod. WHAT I BUILD ✅ Voice AI agents: Retell, Vapi, OpenAI Realtime API, Twilio, ElevenLabs, WhatsApp. Inbound and outbound, qualification, booking, intake, human handoff. ✅ RAG and document intelligence: hybrid retrieval, chunking that fits your documents, metadata filters, evals so accuracy is measured rather than assumed. pgvector, Pinecone, Qdrant, ChromaDB. ✅ Workflow and backend automation: n8n, Make, FastAPI services, webhooks, CRM integrations (HubSpot, GoHighLevel), Postgres/Supabase, Redis queues, dashboards. ✅ LLM infrastructure: self-hosted vLLM, private deployments, cost and latency tuning, observability. Stack: Python, FastAPI, LangChain, LangGraph, Claude, OpenAI, Retell, Deepgram, Twilio, n8n, Qdrant, PostgreSQL, Supabase, Redis, Azure, AWS, Docker. WHERE THESE PROJECTS BREAK Voice: the hard part is not getting the agent to talk. It is capturing messy real-world answers as structured fields, so your team gets clean data instead of transcripts to read. RAG: retrieval itself is rarely the problem. Weak chunking, missing metadata filters and answers that sound right but are not grounded in the right document are. Every RAG build I ship comes with evals. Automation: the workflow is the easy part. Preventing duplicates, retrying failed API calls and making sure a human can see what happened when something breaks is the real work. HOW WE START 1. Book a 30-minute consultation or message me with one workflow to automate, or one agent that isn't working. 2. You get a written scope with milestones and what "done" means for each one. 3. I ship in small milestones you can test yourself: a number you can call, a query you can run. 4. Every build ends with logs, evals and a handover doc so your team o