I'm an AI Engineer with 8 years of production software experience, focused on building Large Language Model (LLM) applications — the systems companies actually ship, not demos.
I build Retrieval-Augmented Generation (RAG) pipelines, LLM agents with tool calling, Model Context Protocol (MCP) integrations, structured-output extraction systems, and computer-vision pipelines (object detection, OCR, face recognition). My work emphasizes what most AI tutorials skip — reliability, multi-tenant security, evaluation, cost control, and graceful failure.
That discipline comes from years building fintech platforms (processing $1M+ in transactions), high-traffic systems (100,000+ daily active users, an API tuned to 1M+ requests), and Layer 1 blockchain infrastructure. I already know how to ship secure, scalable systems — I apply that same rigor to the AI layer.
The AI is learnable in months; the engineering judgment took 8 years. I bring both.
🌍 Open to remote roles — full overlap with Gulf hours, 3–4 hours with UK & Europe every afternoon, US mornings covered. Contract or full-time.
On code visibility: production work for client engagements lives in client-owned repositories. The AI repos below are public architecture write-ups — design decisions, failure modes, and what I'd do differently — rather than client code.
Production RAG API built with Python, FastAPI, ChromaDB, and Ollama. Token streaming via Server-Sent Events, tenant-isolated retrieval enforced at query time rather than post-filtered, cost/latency observability middleware, and provider-failure handling. The write-up covers how the isolation property is actually tested — near-identical embeddings across tenants, so the vector store has every reason to confuse them.
Both halves of a Model Context Protocol integration — a server exposing tools and resources, and a client that discovers and calls them at runtime, wired into a local-model agent loop. Building both sides is what surfaces the real design questions: tool descriptions as prompts, server-side validation, and errors written for a model to act on rather than a human to read.
An LLM agent with a think → act → observe tool-calling loop, step-limit safety caps, and schema validation at the tool boundary. Implemented in raw orchestration and rebuilt in LangGraph — with an honest account of what the framework bought and what it cost.
LLM system that classifies priority, extracts structured fields, and drafts replies from raw tickets. The engineering is in the degradation path: needs_human as a first-class outcome rather than an error state, because a confident wrong classification is worse than an explicit "I don't know."
Computer-vision pipeline integrating YOLO for shelf and product detection, OCR for on-shelf text extraction, and ArcFace for face recognition. Where multi-model pipelines actually break: error compounds multiplicatively, and detection crops are hostile input to OCR.
Saudi Arabia e-invoicing compliance implementation (Fatoora, XML generation, certificate signing) — direct experience with Kingdom of Saudi Arabia regulatory requirements. Open-sourced and in use by other teams (⭐ 13 · 🍴 7).
- AI Systems: RAG pipelines, agents, and MCP integrations built with production concerns — evaluation harnesses (retrieval hit-rate + faithfulness), multi-provider abstraction (Anthropic Claude, OpenAI, Ollama), and defensive structured-output parsing.
- Performance at Scale: Migrated a high-traffic Node.js service serving 1M users to Go, resolving memory issues that had been managed with scheduled restarts.
- Optimized API Performance: Refined a blockchain art API to handle 1M+ requests through careful server-resource management.
- High-Traffic Systems: Scaled a CMS platform to 100,000+ daily active users; cut operational costs 40% moving to serverless on AWS Lambda.
- Technical Leadership: Led projects to CMMI 2.0 certification; optimized startup times to 10 seconds; elevated an app rating from 4.3 ★ to 4.9 ★.
- Fintech: Delivered secure transaction, authentication, and fraud-prevention features for a fintech app processing $1M+ in its first quarter.
| Category | Technologies |
|---|---|
| AI / LLM |       |
| AI Infra |     |
| Computer Vision |    |
| Languages |       |
| Databases |     |
| Cloud & DevOps |       |
| Blockchain |    |
Core engineer on a live Layer 1 chain and the full product surface around it — the backend and protocol depth that underpins my production engineering discipline:
- ⛓️ Elysium Chain — Layer 1 blockchain (Substrate + Frontier EVM compatibility layer) for AI, gaming, and metaverse projects. Aura consensus configuration, validator-set and node-authorization pallets, load testing and gas-limit tuning.
- 🔍 Elysium Scan — Real-time blockchain explorer for transactions, blocks, and accounts.
- 💼 Elysium CloudWallet — Web3 wallet for managing digital assets.
- 🎨 Elysium AI NFT Studio — No-code AI NFT generation and launch platform.
- ⚙️ PoS Consensus in Go — Proof-of-stake consensus from scratch: stake-weighted leader selection, block proposal, cross-node validation.
- 🛒 Shopify Apps — Cartly, Sales Pop-up, and Wheelify — used by thousands of merchants.
- Email: se.sulaman.khan@gmail.com
- LinkedIn: sulaman-khan
- WhatsApp: Click to chat