Arij I.

I build production RAG systems and LLM agents in Python. Most AI projects stall in the same two places: retrieval returns the wrong chunks, and nobody can prove the system got better after a change. I fix both, with evals and tracing from day one. WHAT I BUILD RAG over your documents Ingestion and parsing for messy real files: PDFs with tables, scanned documents, Confluence exports, support tickets. Chunking that respects document structure instead of cutting blindly every 500 characters. Hybrid search combining keyword and dense vectors with reciprocal rank fusion, then reranking. Answers that cite their sources and refuse when the corpus does not support them, because a confident wrong answer costs you more than an honest "I don't know". LLM agents and workflows Multi-step agents in LangChain and LangGraph: tool calling under explicit permissions, conversation state that survives restarts, retry and idempotency around every external call, and human-in-the-loop approval gates on anything touching money, personal data, or a customer-visible action. Two years on agents handling real inbound volume means I build for the failure modes, not for the demo. Evaluation and observability This is the part most freelancers skip, and it is why projects stall at 80 percent. I build you a golden question set, score it with RAGAS plus metrics specific to your domain, trace every run in Langfuse, and wire a CI gate so a prompt change cannot silently regress quality. You get a number that tells you whether the system improved, not a feeling. The backend around all of it FastAPI or Django services with streaming responses, background jobs in Celery, PostgreSQL with pgvector, Redis, Docker, and explicit cost and latency budgets per request. Nine years of production backend work is the reason the AI part stays up once real traffic arrives. HOW I WORK - Scoped in writing before we start. You get a plan with milestones, not an open-ended hourly drip. - Something you can click by the end of week one, even if it is rough. You should never be three weeks in wondering what you bought. - Measurable acceptance criteria. We agree what "working" means numerically before I build it. - Updates on a schedule, and I tell you when something is not working before you have to ask me. A GOOD FIT IF - You have a document corpus and users who keep getting bad answers out of it. - You have an AI feature that works in testing and falls over in production. - You have a chatbot or agent that is inaccurate and nobody can explain why. - You need a Python backend built properly around an LLM product. NOT A GOOD FIT IF - You need model training or fine-tuning from scratch. I use, integrate and evaluate models. I do not train them. - The project is computer vision. - The brief is "build me an AI app" with no defined problem yet. Book a consultation instead and we will scope it properly first. STACK Python, FastAPI, Django, LangChain, LangGraph, OpenAI, Anthropic, pgvector, Chroma, Qdrant, RAGAS, Langfuse, Celery, Redis, PostgreSQL, Docker, AWS, n8n, BACKGROUND Nine years building and maintaining production backends, the last two full time on LLM agents for recruitment and customer support. Several of those years were spent on healthcare systems, so HIPAA constraints, PHI handling and audit trails are familiar ground rather than a surprise discovered halfway through a build. The completed contracts on this profile are web and automation work, and they are the reason I hold a 100% Job Success Score across 1,659 hours and 70 jobs. I scope honestly, I communicate on a schedule, and I do not disappear. The technology here is new. The way I run a project is not. FIRST STEP Send me your corpus size, your document types, and one example of an answer that came back wrong. I will tell you whether the problem is retrieval, chunking or the prompt before you hire me.