Mohammed Usmani · AI Engineer, MumbaiThree models.
One GPU.
In production.

I deploy open-weight LLMs onto GPUs and keep them serving. This drawing is the stack running our production traffic today, on one AMD MI325X. I also build the agents, retrieval and voice systems that call it.

Control plane · Godeploys models onto GPU nodesfit checkrefuses a model that won't fit,tears down on the 2nd crashGateway · own model firstthen Claude, then OpenAI60 scircuit breaker before fallingback to hosted APIsvLLM 0.32 · priority schedulinguser calls 0 · background jobs 10~11×background dossier builds,350 → 3,900 an hourQwen 27B · FP8Qwen 9Beheadroom210 / 256 GiBin use · ~29 full 32k-contextrequests at once210 GiB in useAMD MI325X · 256 GiB HBM · e = embedding model
DeployMachine production stack, top to bottom: a Go control plane that deploys models and refuses ones that will not fit; a gateway that tries our own model first, then Claude, then OpenAI, with a 60 second circuit breaker; vLLM 0.32 with priority scheduling, user calls at priority 0 and background jobs at 10, which took background dossier builds from about 350 to 3,900 an hour; and one AMD MI325X with 256 GiB, 210 GiB in use, split 60% to Qwen 27B with FP8 KV cache, 20% to a 9B model and 5% to an embedding model, serving about 29 full 32k-context requests at once.
GPU memory in use
210 / 256 GiB
full 32k-context requests at once
~29
background builds after priority scheduling
350 → 3,900/h
breaker before falling back to hosted APIs
60 s
Drawn byMohammed UsmaniAI Engineer · Mumbai · TechNsure IT Hub
Sheet01 of 15
Rev2026-11
StatusOpen to inference roles
Issuerésumé.pdfContactmohammedusmani2005@gmail.com

Notes · ask this drawing

Sheet index

15 drawings · billed by weight
SHEET 01DeployMachineDeploys open-weight LLMs onto GPUs in the customer's own cloud, behind an OpenAI-compatible endpoint.drawn above · full case study →
Other drawings in the set
SheetTitle
05HookRelayWebhook delivery that survives crashes, duplicates and receivers that stay down for days.
06FS Preparation AgentFinancial statements where the LLM points at cells and arithmetic decides every link.
07Saffron & SmokeVoice ordering where the agent can only speak prices its tools returned.
08Agent Orchestration LayerTurned a single-user CLI agent runtime into a multi-tenant service without forking it.
09GarageCrash-recoverable meeting note-taker that picks single-pass or map-reduce by transcript length.
10JobTrackA Gemini insights layer that walks six models down when the free quota runs out.
11SoulmateA paid matchmaking app where double-credits and double-charges are ruled out by schema constraints.
12CityFixA civic issue-reporting app, 3rd Prize at GNEC 2024; I built the backend and schema.
13SolidBoardA SolidJS Kanban board with two runtime dependencies and dense-integer card ordering.
14Finance APIA typed-error REST API for expenses, with two dates per entry and a duplicate guard.
15GameWebA browser gaming portal whose ETL merges two game feeds idempotently into one schema.
All 25 systems, including the 10 production services without their own sheet →
Rev
Description
A

Associate Software Developer (AI Engineering)

TechNsure IT Hub Pvt. Ltd. · Remote · Jan 2026 – Present

  1. A.1Measured 956 production recommendations, then replaced a useless score threshold with a judge that may say nothing fits: forced product pitches fell from 100% to 8%.
  2. A.2Built an agentic sales copilot over seven context layers that skips raw RAG when its relationship model is fresh, and reranks by ID only so it can't mislabel products.
  3. A.3Mirrored a 9-type catalog into Qdrant through a transactional outbox and HMAC webhooks, re-embedding only items whose text actually changed.
  4. A.4Moved Facebook and Instagram scraping onto users' devices, so no session is ever replayed from two IPs.

Bill of materials

79 parts · used in shipped work

Languagesqty 7

LLMs & agentsqty 19

Retrieval & RAGqty 9

Voice & realtimeqty 8

Backendqty 12

Dataqty 9

ML & visionqty 3

Cloud, infra & frontendqty 12

Top 17 of 5000+AI For Good Hackathon2025
3rd PrizeGNEC International Hackathon2024
3rd PrizeCodeathon DSA Hackathon2024

General notes

answers a recruiter usually needs

Who is Mohammed Usmani?

Mohammed Usmani is an AI Engineer based in Mumbai, India, currently an Associate Software Developer (AI Engineering) at TechNsure IT Hub, working remotely. He builds production agentic AI systems — multi-agent orchestration, retrieval pipelines, realtime voice, and the multi-tenant backends behind them — and has shipped 15+ of them across sales, accounting and workspace products.

What AI systems has he actually built?

Four stand out. A financial-statement engine whose Cross-Reference Graph links every statement line to its note and ledger account, confirmed by arithmetic tie-out rather than model judgement. Garage Connect, an AI sales platform where an opportunity judge measured on 956 production recommendations cut forced product pitches from 100% to 8%, alongside an agentic copilot, a cited relationship model and on-device social capture. Luca, a ~199,000-line AI accounting platform for Indian Chartered Accountants running on five LLM providers. And EyesAI, an autonomous Android agent that operates a phone end-to-end for blind and low-vision users.

What is his experience with RAG and retrieval engineering?

He built a seven-layer prompt assembly for a sales copilot that cut estimated worst-case input from roughly 96,000 to 2,000 tokens per turn, with six Qdrant collections across the platform. The retrieval goes beyond nearest-neighbour search: contextual retrieval that prepends an LLM-written blurb before embedding, HyDE expansion for terse queries, a 14-day recency decay, LLM reranking, and reciprocal rank fusion of dense and text search in a webinar assistant. He has also worked with pgvector and FAISS.

Has he built realtime voice AI?

Yes. He built a live in-meeting copilot that captures the host microphone and remote participants as two separate streams, downsamples 48kHz to 16kHz PCM16 in 250-millisecond batches, and gives each stream its own Deepgram socket — so speaker attribution comes from the transport rather than a diarization model. He also built a voice-to-voice restaurant ordering system on Pipecat and Gemini Live over a full-duplex WebSocket, and telephony voice agents over Telnyx SIP.

What is the Cross-Reference Graph?

It is the core data structure of the financial-statement engine: 9 node kinds and 6 edge kinds modelling how a set of financial statements actually hangs together. Links are made by deterministic resolvers and then confirmed by arithmetic tie-out, and every edge records how it was created and whether the numbers reconcile. Where an LLM assists, it only extracts and points at cells — the pass/fail verdict comes from the same deterministic arithmetic, so no reconciliation passes without a verified tie-out.

Which programming languages and tools does he use?

Primarily Python with FastAPI and TypeScript with Node, Express and NestJS. Day to day that means LangGraph, the Model Context Protocol, OpenAI, Anthropic Claude and Gemini, Deepgram and Voxtral for speech, Qdrant and pgvector for retrieval, Celery and BullMQ for background work, PostgreSQL, MongoDB and Redis for storage, and Docker with GitHub Actions deploying to GCP, AWS and DigitalOcean.

Is he available for hire, and where is he based?

Yes — he is open to AI/ML and backend engineering roles at product-focused startups. He is based in Mumbai, India, and can be reached at mohammedusmani2005@gmail.com or on LinkedIn.

Need someone to run your models?

I'm looking for AI inference and AI engineering roles. The fastest way to reach me is email; the résumé has everything on one page.

Drawn byMohammed UsmaniAI Engineer · Mumbai · TechNsure IT Hub
Sheet15 of 15
Rev2026-11
StatusOpen to inference roles
Issuerésumé.pdfContactmohammedusmani2005@gmail.com