Work

Lead AI Engineer (full-time) at ILM AI — UK-based, remote — on AI learning and assessment systems. Consulting with Luminary AI on agentic workflows, the Nova workflow platform, and the Prism app builder, and with Scope AI on Suuplies. Part-time and past work spans edtech at ACS Future School, AI/automation at Priyo, and research at CivilGPT (BUET). Hackathons, CTF placements, and programming contest results are on the Awards page.

Full Time

ILM AI — Lead AI Engineer (UK, remote)

Jan 2026 – Present · ilmai.co.uk

Stack: Next.js, FastAPI, PostgreSQL, Pgvector, Qdrant, S3, Redis, Tailwind, shadcn/ui, Gemini, OpenAI.

Focus:

  • Question intelligence: AI-powered question recommendation with vector embeddings, plus student data analytics to inform practice and personalization.
  • QGenAI: a fully autonomous question-generation agent that produces questions across 17 structured question types from large volumes of unstructured source material—improving generation quality and cutting 90%+ of manual authoring cost.

Consulting

Luminary AI — Consultant (Virginia, USA, remote)

Oct 2025 – Present · luminary-ai.com

Stack: Django, DRF, Django Channels, Redis, Apache Kafka, Django-Q2, PostgreSQL, S3, Keycloak, Kubernetes — plus Nova / Prism surfaces on Next.js, BullMQ, E2B sandboxes, Fly.io deploys, Prisma, SSE (Redis pub/sub progress), Anthropic agent loops, and a FastAPI payment path (Stripe, Kafka-driven credits, ledger) where applicable.

Products and contributions:

  • Nova — no-code AI agent workflow builder (n8n-style) with 30+ node types: conditional routing, loops, Gmail, Google Drive, Sheets, Docs, Telegram, vector search, web search, and more — on a microservice-oriented stack (Kubernetes, Kafka messaging, Django-Q cluster background work, Redis caching, multi-process runtime).
  • Prism — low-code web app builder (v0 / Lovable-style): E2B for live coding and preview, Fly.io for long-lived deployments, S3 for code and assets, BullMQ workers driving multiple background agents across projects.
  • Broader platform: backend APIs and WebSocket flows for workflow execution, chat, and document/RAG; async hand-offs (Redis ↔ Kafka producers/consumers); agentic tool-call and workflow-resume patterns; OIDC / API-key auth; observability and production operations on Kubernetes.

Scope AI — Full Stack Developer (USA, remote)

Jan 2025 – May 2025 · Suuplies

Product: Suuplies — procurement- and contracting-adjacent workspace: SAM.gov data, uploaded documents, and LLM-assisted flows with retrieval grounded in your corpus.

Stack: Next.js, TypeScript, React, FastAPI, Python, Clerk, Tailwind, shadcn/ui, LangChain, PostgreSQL, pgvector, Qdrant, Redis, Azure Blob, Drizzle ORM, Nginx, WebSockets (Yjs).

  • Refactored and maintained production full-stack codebase for Suuplies (dashboard, projects, analytics, uploads).
  • Document AI / RAG: PDF and XLSX ingestion, chunking, embeddings, vector search and document scoring (Llama Parse, unstructured PDF parsing, pandas-style spreadsheet paths).
  • SAM.gov: typed Python client (SAMClient) for federal opportunity search and detail—filters (NAICS, PSC, set-aside, geography, dates), reference data, and resource fetch.
  • Integrated 50+ external APIs and LLM-assisted workflows; RBAC and communication integrations.
  • Deploy / infra: Azure VM deployment notes, Nginx reverse proxy and TLS, tmux-based process layout, separate processes for API, frontend, and optional Yjs collaborative WebSocket server.

Part Time

ACS Future School — Backend Developer

Nov 2024 – Aug 2025 · Dhaka, Bangladesh (hybrid) · acsfutureschool.com

ACS Future School is a major Bangladesh online-learning platform—structured academic programs (grades 6–10), live and interactive classes, assessments and leaderboards, materials and community, plus Curiosity AI for curriculum-grounded Q&A over the teaching corpus.

Stack: Node.js, Express.js, Next.js, LangGraph, WebSocket, bKash and payment-adjacent flows, OTP and notification services, Teachmint → BunnyCDN media pipeline.

  • Curiosity AI & RAG: Owned Curiosity from first production release through a tuned retrieval stack over ~12.5 GB / 1000+ PDF/DOCX—chunking, retrieval, and answer quality under real student load (5k+ messages in the opening days). Built the evaluation story, not just the model: subject evals where Curiosity led ChatGPT by roughly 10–15% in places, plus a 16k-question benchmark from production-style data with about ~10% overall lift—end-to-end RAG, measurement, and iteration.
  • Live learning at scale: Led backend design for the live-class surface—data model, REST and WebSocket APIs, deploy path, and deadline-driven releases—shipping chat that held 5k+ concurrent users, inside a product context trending toward 80k+ users across classes, commerce, and analytics-shaped services.
  • Reliability & incidents: Primary responder (often lead or co-lead) on high-severity production events—DoS against OTP/auth, EC2 capacity pressure, database compromise, disk exhaustion, and live-class outages—owning containment, recovery, and durable follow-up (limits, monitoring, operational hardening).

CivilGPT (BUET) — Research Assistant

Jul 2024 – Nov 2024

Stack: FastAPI, SQLAlchemy, LangChain, Next.js/React.

  • Domain-specific RAG for civil engineering across retrieval, parsing, and answer generation.