Mohammed Usmani

resume

Mohammed Usmani

mohammedusmani2005@gmail.com+91 86899 08731mohammedusmani.meLinkedInGitHub

Applied AI engineer building production agentic, retrieval and realtime-voice systems on the multi-tenant backends that carry them. Shipped 15+ AI features and services across sales, accounting and workspace products using FastAPI, LangGraph, Qdrant/pgvector, OpenAI/Anthropic/Gemini and Deepgram, and a Go control plane that deploys open-weight LLMs on GPUs in customers' own AWS and DigitalOcean accounts.

Experience

Associate Software Developer (AI Engineering)

– Present

TechNsure IT Hub Pvt. Ltd.

Remote

  • Own the AI layer of NetworkChains, a live relationship-selling platform: social and messaging ingestion, identity resolution, a relationship model, signal detection, product opportunities and an agentic sales copilot.
  • Measured 956 production recommendations and found the relevance score couldn't separate good from bad matches; replaced it with a per-item LLM judge allowed to say nothing fits, plus a separate writer, cutting forced product pitches from 100% to 8% and hedging from 83% to 0% on replayed cases.
  • Built EarnGPT, a 5-iteration tool-calling copilot over a 7-layer context stack (relationship replica, memory slots, contextual retrieval with HyDE, document RAG) that skips raw RAG when the replica is fresh (estimated ~96K → ~2K tokens/turn); the reranker returns IDs only, so it can't mislabel products.
  • Designed and built DeployMachine (Go, PostgreSQL, Next.js), which deploys open-weight LLMs with vLLM on NVIDIA and AMD GPUs in the customer's own cloud (DigitalOcean; AWS via a cross-account role), with a fit check before any GPU is bought. It now serves three models (27B FP8, 9B, embeddings) on one AMD MI325X to our production sales-assistant backend, falling back to Claude then OpenAI; an A/B on 30 real contacts kept extraction on the 27B (the 9B broke the schema 3/30).
  • Synthesised each relationship from six sources with every claim cited and verified by exact match or ≥0.8 similarity; identity resolution auto-merges only on strong system-derived keys, so uploaded CSV ids can't poison identities, and merges are reversible.
  • Mirrored the Garage catalog (9 item types) into Qdrant through a transactional outbox and HMAC-signed webhooks, skipping re-embedding when an item's text hash is unchanged, with an hourly reconcile and a daily orphan sweep as backstops.
  • Moved Facebook and Instagram scraping onto users' devices (Electron over CDP, hidden mobile WebViews replaying the platform's own GraphQL) so no session is replayed from two IPs; LinkedIn and WhatsApp run server-side behind per-user geo-matched proxies with an API-contract canary.
  • Shipped a live call copilot (one Deepgram stream per audio source, LangGraph), a crash-recoverable LiveKit note-taker, and turned OpenClaw into a multi-tenant agent service: 265 tools, 57 integrations, 36 OAuth providers, a 40-tool MCP server.

Projects

voiceAgent

— AI Phone Screener
  • Python
  • LiveKit
  • LangGraph
  • FastAPI
  • PostgreSQL
  • Turns a job description into screening criteria a recruiter approves, calls candidates over SIP with a LiveKit voice agent, and scores each answer with the candidate's own words as evidence; rejections wait for a recruiter. Logs every turn-taking decision: median reply 2.44 s → 1.22 s at the agent, stopped for 15/15 interruptions on 10 test calls, and dropped a faster model after the log showed it returning empty replies.

Engram

— Shared Memory for AI Assistants
  • Python
  • FastAPI
  • MCP
  • PostgreSQL
  • pgvector
  • mem0
  • Self-hosted memory shared by Claude Code, ChatGPT and other assistants over a 22-tool MCP server with OAuth 2.1; every write must name the agent and when the event happened. One LLM call extracts typed events, facts and procedures and supersedes changed facts; recall fuses 8 channels with rank fusion into a token-budgeted brief. Picked the extraction model by eval (gpt-oss-120b, 126/126).

Technical Skills

AI & Agents
  • Agentic tool-calling ·
  • RAG ·
  • LoRA fine-tuning (Unsloth/PEFT, distillation) ·
  • vLLM ·
  • llama.cpp (GGUF) ·
  • hybrid retrieval (RRF, HyDE, rerank) ·
  • LangGraph ·
  • LangChain ·
  • MCP ·
  • OpenAI ·
  • Anthropic Claude ·
  • Gemini ·
  • Vertex AI ·
  • Qdrant ·
  • pgvector ·
  • Deepgram ·
  • Pipecat ·
  • LiveKit
Backend
  • Python (FastAPI) ·
  • Go ·
  • TypeScript (Node.js, Express, NestJS) ·
  • Celery ·
  • BullMQ ·
  • WebSockets ·
  • REST ·
  • GraphQL ·
  • MCP ·
  • SSE
Databases
  • PostgreSQL ·
  • MongoDB ·
  • Redis ·
  • Supabase ·
  • Neo4j
DevOps
  • Docker ·
  • GitHub Actions ·
  • Nginx ·
  • AWS (EC2, S3, Route 53, CloudFormation, Lambda, RDS) ·
  • GCP ·
  • DigitalOcean ·
  • Railway ·
  • Linux ·
  • CI/CD
Frontend
  • React 18 ·
  • Next.js 15 ·
  • TailwindCSS

Education

Viva Institute of Technology

Mumbai, India

B.E. Computer Science (AIML) | CGPA: 8.27/10

Achievements

  • Top 17 Nationally — AI For Good Hackathon, Blend360

  • 3rd Prize, Codeathon — Vidyavardhini College of Engineering

  • 3rd Prize — GNEC International Hackathon – Global NGO Executive Community