AI Engineer. I build production systems in TypeScript and put models inside them.
Currently the engineer on TrustCart ERP — a multi-tenant platform serving six retail storefronts from a single backend. NestJS and PostgreSQL, TypeORM migrations, Redis and Bull for queues, Socket.IO for realtime, Next.js on the front. 41 modules, 166 entities, 256 pages. I own the schema, the migration discipline, the backup pipeline and the deploys — around 1,300 commits and 620 pull requests so far.
Separately I work on Bengali-language NLP, and on projects that let me push into unfamiliar ground.
Sylhet, Bangladesh · open to remote
| Project | What it is |
|---|---|
| Ankur · live | Turns Bengali and English documents into source-grounded adaptive assessments. Every generated item must cite validated source segment IDs before it can be displayed — the source boundary is enforced, not suggested. |
| Caissa | A local-first chess application with a bounded engine and guided post-game review. No accounts, no server, no analytics; everything runs on your device. 103 test files, CI, GPL-3.0. |
| AirScript | Touchless air-writing. MediaPipe hand tracking in the browser, CNN inference and NLP correction behind FastAPI. |
| SelfDiary | Cross-platform diary — Next.js web and Expo mobile, offline-first with automatic conflict resolution on reconnect. |
| Multi-tool AI Agent | An agent that routes across multiple external tools, in Python. |
| Gym Sales OS | Lead qualification, trial booking and retention automation. n8n, OpenRouter, Google Sheets, Gmail and Cloudflare Workers. |
Bengali is badly under-served in NLP and in security research, so that is where my research time goes. Four studies are in progress or under review:
- Hallucination detection in Bengali — how detectors trained on synthetic data hold up when they meet natural text.
- Benchmark integrity for Bangla text classification — auditing public datasets for train/test leakage and shortcut learning, and what that means for reported scores.
- Cross-lingual behaviour of code models — whether prompting in English, Bangla or Banglish changes how often a model invents package dependencies, and the supply-chain exposure that creates.
- Password strength meters on Bangla and Banglish input — where meters return dangerously confident scores for passwords they cannot actually reason about.
Repositories stay private while the work is under review.
Daily — TypeScript · NestJS · Next.js · React · PostgreSQL · MongoDB · Node/Express
Also — Python (FastAPI, TensorFlow) · Kotlin · Redis · Docker · n8n
What I actually care about — migrations that do not diverge, backups someone has tested, tests that catch the regression, and a README the next person can operate from.