Mahdi Bani

I'm an AI engineer with a backend foundation. I build and operate production LLM systems end-to-end — document AI, LLM agents, RAG, and automation, and I focus on the part most people skip: making AI dependable in production. Reliability, cost, guardrails, testing, observability, the unglamorous work that decides whether an AI feature survives contact with real users.

Most of my experience is ownership at the deep end. I solo-built and run a multi-tenant AI platform as the entire engineering function for a non-technical company: LLM agents on WhatsApp and inbound voice (Claude tool-use, Deepgram, ElevenLabs) with deterministic guardrails, two-way Microsoft 365 sync, real-time WebSockets, and a Dockerized FastAPI + Celery backend — 20 API services and 18 data models across the stack. In parallel I built and own a production document-AI pipeline (54 document types, GPT-4o vision + DSPy) where I cut LLM inference cost 5.7× while holding 88% F1, backed by 438+ automated tests in CI and a cost-tracking dashboard that made real AI spend observable.

I'm also the person who keeps AI running when the ground shifts: I've contained a DSPy major-version break, a litellm supply-chain compromise (in under an hour), and breaking voice/Meta API migrations before any of it reached users. Earlier I built a crypto-exchange end-to-end (real-time pricing, KYC, fiat-to-crypto), fine-tuned LLMs for client work, and contributed real open source — including 2 merged PRs to the Bittensor core SDK, one fixing a 2.4 GB memory leak.

Stack: Python, FastAPI, SQLAlchemy, PostgreSQL, Redis, Celery · React, Next.js, TypeScript · DSPy, LLM agents & tool-use, RAG & embeddings, fine-tuning · Claude, GPT-4o, Llama/Gemma (local) · Docker, CI/CD, Azure, multi-tenant RBAC · Deepgram, ElevenLabs, Pipecat (voice).