I am a final-year undergraduate in Computer Science and Engineering at Bangladesh University of Engineering and Technology (BUET). My interests are primarily in Distributed Systems, ML Systems and Applied AI - building software that stays correct and cheap under load.
I have previously worked part-time as an AI Engineer at CreateBud, Inc., a Research Assistant at Sysmodeler.ai on Model-Based Systems Engineering and AI-based SysML generation, and a Backend Engineer at Systemica Institut. On the side, I work as a software consultant and freelance for small and mid-size businesses, and contribute to open source - including a merged PR to Plone.
I’ve competed a few times in my undergraduate years - winning the IEEE ICIP 2025 VIP Cup and the AI Engineering Hackathon (poridhi.io × Brain Station 23). I have tutored university and college students for three years, chaired the technical committee at IEEE CS BUET SBC, and now serve as its Vice Chairperson (Technical). Outside engineering I travel, lift and listen to far too much music - I write about all of it. If you want to talk, just shoot me an email!
RAG systems can give the right answer while citing a source they never used - a failure called post-rationalization. With Md. Shariful Islam Bhuyan, I added a missing control to an existing methodology for measuring this, then used it to test whether training search agents with RLVR makes citations more honest. It doesn't: across four QA datasets, RLVR-trained agents post-rationalized at close to the base model's rate.
Team NeuronX won the IEEE ICIP 2025 VIP Cup among 40+ international teams, building fusion pipelines for real-time multimodal (RGB + infrared) drone detection, tracking and payload classification under fog, occlusion and thermal-contrast gaps. Supervised by Prof. Dr. Shamsuzzoha Bayzid, with graduate tutor Asib Rahman.
Team NeuronX - me, Abrar Zahin Raihan, Sadik Mahmud Shakshor, Sadman Sakib and Ruwad Naswan - won the IEEE ICIP 2025 VIP Cup among 40+ international teams, building fusion pipelines for real-time multimodal (RGB + infrared) drone detection, tracking and payload classification under fog, occlusion and thermal-contrast gaps. Supervised by Prof. Dr. Shamsuzzoha Bayzid, with graduate tutor Asib Rahman. I presented our solution live at ICIP 2025 in Anchorage, Alaska.
Team BUET XFACTOR won the AI Engineering Hackathon (poridhi.io × Brain Station 23) among 86 teams and 300+ participants, building a multilingual intent-based search system with a 500ms average response time - without a GPU, using only half the provided compute (10 of 20 VMs) via in-memory llama.cpp inference and 4-bit quantization.
Top 10 finalist among 1300+ participants at Therap JavaFest 2024 with Mustafa Muhaimin, building a Spring Boot mentorship platform - session booking, real-time chat, payment integration, and AI-powered mentor matching via Recombee and ChatGPT. Invited to present at the grand finale at Therap BD Limited headquarters.
A production RAG backend for an ed-tech platform covering Bangladesh's NCTB higher-secondary curriculum (Physics, Chemistry, Math, Biology). Three idempotent Celery pipelines share one corpus: one OCRs textbooks, chunks them, classifies them against a 1,075-topic taxonomy and embeds them into Qdrant; a second retrieves grounding chunks and generates new bilingual (Bangla/English) MCQs, with figures the model draws as sanitized SVG; a third ingests scanned past papers end to end - extraction, Bangla-aware answer-key binding, retrieval-grounded answer resolution, and bilingual explanations. Every MCQ from either path clears a semantic-dedup check and a cross-model LLM judge before it's stored; on a 96-MCQ benchmark, 90/96 came out usable at a judge mean of 8.2/10, for about $0.02 per usable question. Built a provider-agnostic LLM layer (OpenAI/Anthropic) that treats model output as untrusted input - structured-output schemas, per-field validators and graceful degradation at every layer - plus Celery delivery guarantees (acks_late, reject_on_worker_lost) so a crashed worker never loses or duplicates a job. Stack: FastAPI, Celery, PostgreSQL, Qdrant, Docker, DigitalOcean behind Caddy TLS. Thorough unit and e2e test coverage, with GitHub Actions CI (lint, e2e against live APIs, migration-drift check, dependency audit) and automatic deploy on every passing push. Built through my software consultancy work - proprietary, so code isn't shareable.
Systems, backend and infrastructure lead for Netra News' live election monitoring platform covering Bangladesh's 2026 national election - 300 constituencies, ~43,000 polling stations. Built as a static Next.js site using Incremental Static Regeneration, triggered by authenticated webhooks on each result entry, backed by a FastAPI monolith and a PostgreSQL primary/replica pair with two-layer connection pooling. Deployed on a DigitalOcean Kubernetes cluster behind Cloudflare (WAF, DDoS protection, Tunnels in place of a load balancer). Served roughly 8 million requests at 100% uptime for about $335 in infrastructure cost for the month.
A full compiler for a subset of C, built for CSE310: Compiler Design. Complete pipeline - Flex lexical analysis, ANTLR4 parsing, a custom C++ symbol table, semantic analysis and intermediate-code generation via ANTLR visitors - down to x86 assembly that runs on EMU8086 and DOSBox.
A buy/sell marketplace for new and used vehicles, built for a client with a friend - I owned the entire backend. Django REST Framework, React and TypeScript, JWT auth, hosted on DigitalOcean; continuous client revisions were a good lesson in production-grade scaling and client handling.
A platform for organizing treasure hunts - teams, puzzles, real-time leaderboards and announcements. Built with Django REST Framework, React and TypeScript; my first SaaS product and my first app used by many people concurrently, which meant learning production-level debugging under real load.
An online store for a local clothing brand with payment integration (SSLCommerz) and three-layer authorization. Django REST Framework, React, JWT auth, tailwindCSS - the two hardest parts were the payment integration and the multi-layer auth.