Usama M.

I build AI products that make it to production: RAG document intelligence, video AI pipelines on self-hosted LLMs, and AI-first SaaS platforms; hosted or self-hosted, depending on where your data is allowed to live.

$200K+ earned, 6,000+ hours, 100% Job Success, Top Rated Plus.

WHY FOUNDERS AND TEAMS BRING ME IN:

✅ Built the AI behind a lending platform with 1,000+ SBA loans and $30M+ in funded transactions, still running today
✅ Turned decades of broadcast footage into an archive you search by asking a question, on self-hosted GPUs because the data couldn't leave their network
✅ RAG that cites its sources, so you're not fixing invented answers a month in
✅ You work directly with me on architecture and code. When a project needs more hands, I scale with my own team at Arithmiks rather than handing you off

WHAT I BUILD:

📄 1. Document AI, Contract Review and Intelligent Document Processing
Layout-aware parsing and OCR (Docling, AWS Textract), clause-level risk flagging, commercial term extraction and exact-section citations, built on RAG rather than a model answering from memory. Models hallucinate sources, so every claim is checked against the real text and dropped if it can't be verified. Shipped inside a multi-tenant lending platform for SBA lenders.

💬 2. RAG Systems, AI Chatbots and Conversational AI
Production RAG with hybrid dense and sparse embeddings (Qdrant, FastEmbed), reciprocal rank fusion, reranking, and query rewriting so follow-ups still retrieve the right evidence. Multi-document RAG, knowledge-graph retrieval (Cognee), RBAC-scoped retrieval and per-org data isolation. Every answer carries a clickable citation. Provider-agnostic routing across Claude, GPT, Gemini and open-weight models.

🎥 3. Video, Audio and Media AI Pipelines
Speech-to-text (Whisper/ASR) with speaker diarization, subtitles, face detection, chapter segmentation, entity tagging, and semantic search with RAG chat over indexed video libraries. Multilingual, Arabic and English with full RTL support. Orchestrated as DAG workflows on Argo over Kubernetes, each video isolated so one failed stage retries alone.

⚙️ 4. AI Agents and Workflow Automation
Multi-agent pipelines with conversation state, self-repair and idempotent retryable stages. Meeting intelligence bots that join Zoom, Google Meet and BigBlueButton, transcribe live, and generate decision logs, commitments and follow-up emails. Text-to-SQL over Postgres and MySQL with AST-level validation and read-only execution.