AI Engineer & Researcher
RL post-training · Agentic systems · Full-stack AI products
I build language-model systems end to end - the reinforcement learning that trains a policy, the agent architecture that puts it to work, and the product it ships inside.
My research is in RL post-training: reward design, GRPO and its variants, and the custom environments that make verifiable RL possible in domains where no benchmark exists yet. I'm drawn to the part usually skipped - building the simulator, the verifier and the reward signal, then measuring honestly whether RL earned its compute over supervised fine-tuning.
The engineering half is agentic. Memory architectures that let agents carry context across sessions (the subject of my WOA 2025 paper), tool calling and multi-step orchestration, graph and hybrid retrieval, and the adversarial testing that shows where those tool chains break. Around all of it sits the ordinary work that decides whether a model ever reaches a user: inference services, vector search, auth, payments, realtime web and mobile clients, and the CI that ships them. I also care about low-resource language evaluation, particularly Bengali.
CSE, BUET. Based in Dhaka, open to remote research collaboration.
| Layer | Focus |
|---|---|
| Research - RL post-training | GRPO and variants (DAPO, Dr. GRPO, GSPO, RLOO) · multi-signal reward design · physics-grounded and verifiable environments · measuring RL headroom over SFT |
| Agentic systems | Agent memory architectures · tool calling and multi-step orchestration · GraphRAG and hybrid retrieval · structured output and evaluation harnesses · adversarial testing of agent tool chains |
| Full-stack delivery | Inference services and quantised local deployment · vector search · auth, payments and background jobs · Next.js / SvelteKit / Flutter clients · Docker and GitHub Actions |
| Reliability & evaluation | Fault-injection scenarios for incident-resolution benchmarks · adversarial safety datasets for Bengali and other low-resource languages |
MemAgent: A Cache-Inspired Framework for Augmenting Conversational Web Agents with Task-Specific Information N. Sakib, P. Barai, S. I. Parisa, A. Iqbal - WOA 2025, 26th Workshop From Objects to Agents, Trento, Italy. CEUR-WS Vol-4028, Paper 8
An agent memory architecture: a Memory Cache Bank with time-based expiration that decouples information gathering from task execution, so an agent stops re-asking users for details it has already learned. Reduces average conversation turns by 22.4% (5.00 → 3.88) across 150 Mind2Web tasks; a 15-participant study showed a 58% reduction in completion time for recurring tasks.
| Repository | Description | Stack |
|---|---|---|
| dc_ops_environment | Physics-grounded datacenter RL environment - RC thermal networks and continuous multi-objective rewards, built on OpenEnv | OpenEnv · Physics sim |
| dc_ops_training | Teacher-distilled SFT → GRPO pipeline on AMD MI300X. Multi-signal reward (physics, scenario heuristics, anti-looping, format) drove +188% composite reward and a 10× per-step gain on hard multi-fault scenarios | GRPO · TRL · Unsloth · vLLM · ROCm |
| GeoQL-4B | Text-to-OverpassQL via SFT → GRPO on Qwen3-4B, with a documented negative result on reward redundancy in reference-based reward design | GRPO · Qwen3 · vLLM |
| medical-cot-assistant | Clinical chain-of-thought fine-tuning of a 20B model with QAT and LLM-as-a-Judge evaluation; INT4 + GGUF export for local inference | Unsloth · QAT · llama.cpp |
| Repository | Description | Stack |
|---|---|---|
| civilmate-backend | Agentic GraphRAG over building codes - plans multi-step lookups across a code graph, then compares design drawings against site imagery to generate technical logs and blocker alerts | Neo4j · Qdrant · FastAPI |
| InsightAI-python-backend | Microservices AI platform - tool-routed generation of quizzes and flashcards from PDFs and video, enforced structured output, and CLIP-based multimodal product search | LlamaIndex · Qdrant · CLIP · FastAPI |
| Tokkhok-Backend | Personalised RAG chatbot for "Banglish" (romanised Bengali) with a custom transliteration pipeline, few-shot inference and configurable agent personas | FastAPI · Qdrant · PostgreSQL |
| sust-backend | Bebsha AI service layer - RAG product search, description generation, background removal | Flask · RAG |
| Repository | Description | Stack |
|---|---|---|
| bd-prescription-medicine-recognize | ResNet50–CRNN for 78-class handwritten medicine-name recognition on Bangladeshi prescriptions. Hash-grouped StratifiedGroupKFold + SWA ensemble reached 92.53% test accuracy / 0.9234 macro-F1 (+6.44 pt over baseline) | PyTorch · CRNN · SWA · MLflow · Grad-CAM |
| Repository | Description | Stack |
|---|---|---|
| SREGym | Contributed fault-injection scenarios to an incident-resolution benchmark for AI agents - Kafka poison-pill head-of-line blocking, CFS throttling brownout, and oscillating config corruption | Kubernetes · Kafka · Chaos engineering |
| Repository | Description | Stack |
|---|---|---|
| nerdherd2ndrun | Collaborative productivity platform - shared notes, video calls, real-time quizzes, AI assistant | SvelteKit · Firebase |
| coderhub | Developer community platform with blogging, skill-based search, and project management | SvelteKit · Vercel |
| yobofrontend | Frontend for YoboSQL, a text-to-SQL conversational interface | TypeScript · SvelteKit |
| Repository | Description | Stack |
|---|---|---|
| django-init-template | Production-ready Django REST scaffold with PostgreSQL and Supabase auth | Django · DRF · Supabase |
| nodejs-init | Express + TypeScript service scaffold with Helmet, Swagger, and Supabase auth | TypeScript · Express |
Published on Hugging Face @Melikshah.
| Artifact | Description |
|---|---|
| GPT-OSS-20B-Clinical-CoT (GGUF) | 4-bit quantised 20B model fine-tuned for clinical chain-of-thought reasoning, optimised for local inference |
| dc_ops_grpo_lora | GRPO-trained LoRA adapter for the datacenter operations environment |
| qwen3.5-4b-base-blindspots | Adversarial evaluation set probing architectural and logical failure modes of Qwen3.5-4B-Base |
| Shajgoj & General Products | 20,000+ item multimodal datasets for image–text retrieval |
| Prothom Alo News | Large-scale Bengali news corpus for low-resource NLP research |
| Role | Organisation | Focus |
|---|---|---|
| Lecturer, CSE | BRAC University · present | Machine Learning, Operating Systems, Software Engineering, System Analysis & Design |
| Founding Engineer (Backend & AI) | Intellesphere · 2024–25 | Banking RAG with hybrid indexing and Keycloak-secured access, CLIP multimodal search, multimodal civil-engineering automation |
| Founding Engineer (Backend & AI) | Oleyn · 2024 | Bengali ASR with NeMo speaker diarization, agentic CRM and campaign system, legal research assistant |
| Software Engineer (Backend) | Priyo · 2024 | Django support agent at concurrency, warehousing REST APIs, campaign and analytics infrastructure |
| Software Engineer Intern | Yobo · 2024 | Text-to-SQL chat interface on LangChain, FastAPI and SvelteKit |
| Domain | Tools |
|---|---|
| Languages | Python · C++ · Java · TypeScript · JavaScript · SQL |
| Training & Inference | PyTorch · TRL · Unsloth · PEFT/LoRA · vLLM · llama.cpp · ROCm · Transformers · NeMo · scikit-learn · MLflow |
| Reinforcement Learning | GRPO family (DAPO, Dr. GRPO, GSPO, RLOO) · reward design · custom verifiable environments · OpenEnv |
| Agentic Systems | Tool calling · agent memory · multi-step orchestration · LangChain · LlamaIndex · Genkit · RAG · GraphRAG · hybrid retrieval · structured output · eval harnesses |
| Backend & APIs | FastAPI · Django · Spring Boot · Express · Firebase Cloud Functions · Nginx · Keycloak · Stripe |
| Frontend & Mobile | Next.js · React · SvelteKit · Flutter · Tailwind CSS |
| Data & Vector Stores | PostgreSQL · MySQL · Redis · Qdrant · ChromaDB · Neo4j · Supabase · Turso |
| Infrastructure | Docker · Kubernetes · AWS · GCP · Vercel · Modal · GitHub Actions · PostHog |
| Award | Event | Year |
|---|---|---|
| Champion | IUT 11th National ICT Fest - OpenAPI Hackathon | 2024 |
| Champion | CodeCrafters Dev Sprint Hackathon, BUET | 2024 |
| Champion | SUST CSE Carnival Hackathon | 2024 |
| Runner-Up | Gen-Dev Hackathon, Acme AI | 2024 |
| Honourable Mention | Bangladesh Blockchain Olympiad (BCOLBD) · IDSOL World Finalist | 2024 |
| Champion | Cefalo ITverse Project Showcase | 2023 |