I'm a Generative AI Engineer & Team Lead at Joint Venture AI (Betopia Group), where I turn brittle GenAI prototypes into production systems — schema-validated, observable, and tested, not vibes-checked. My work spans document AI (OCR → transformer-based token classification → graph neural networks for structured extraction), LLM inference optimization, and developer tooling that makes GenAI pipelines as rigorously testable as any other piece of software.
Originally trained as an Electrical & Electronic Engineer (Faridpur Engineering College, '25), now based in Dhaka, Bangladesh — building toward AI products for a global market.
🏆 2nd Runner-Up, SEC EEE FEST 2024
A few numbers from the production side of the work:
- 🔭 Leading GenAI systems work at Joint Venture AI (Betopia Group)
- ⚙️ Building Driftgate — a Go-native regression-testing CLI for LLM prompts, solo, nights and weekends
- 📚 Deepening: LLM training infrastructure · Go systems engineering · production MLOps · agentic systems at scale
- 🎯 Open to Generative AI Engineer / AI Systems Tech Lead roles — remote-first, open to relocation
Regression testing for LLM prompts, shipped as a single static binary.
Most prompt changes ship on vibes — "I read the output, looks fine." Driftgate turns that into an actual test suite: define expected behaviors as assertions, run them against every prompt change, catch regressions before they reach production. Written in Go specifically so it compiles to one dependency-free binary — the whole point is that it runs inside air-gapped and regulated enterprise environments where pip install isn't an option. Currently expanding assertion types and CI integrations ahead of a public launch via GoReleaser + Homebrew.
Go CLI Regression Testing Air-Gapped Deployment
A 3-stage extraction system for ACORD 25/101 forms that survives messy real-world scans.
PaddleOCR handles layout-aware text extraction, LayoutLMv3 runs multi-head BIO token classification, and a GAT + GraphSAGE combination reasons over the document spatially — because insurance forms aren't linear text, they're a grid of related fields. Outputs strict, null-safe JSON with hybrid confidence routing (graph ≥ 0.85 → model ≥ 0.75 → rule-based fallback). Cut extraction errors by 35%.
PyTorch LayoutLMv3 PaddleOCR Graph Neural Networks
A systems-design deep dive into adaptive LLM serving — under real constraints, not unlimited cloud budget.
Built entirely on free-tier Kaggle T4 GPUs (16GB VRAM, zero budget), serving the Qwen2.5-Instruct family (0.5B–1.5B, AWQ int4). Development order was eval-first: a frozen eval set and a mock backend shipped before a single GPU cycle was spent — because you can't optimize what you haven't measured yet.
LLM Serving Qwen2.5 Systems Design GPU-Constrained Inference
🗂️ More work — ALPVS, offline vision systems, earlier extraction pipelines
- ALPVS (AI Policy Validation System) — Fully offline, air-gapped policy validation architecture, documented to enterprise standard with a full TDR and CTO-level architecture blueprint.
- Offline OCR/Vision Toolkit — A fully offline Streamlit application built to SOLID principles, with a pluggable model registry for swapping vision backbones without touching the pipeline.
- Early ACORD Extraction Engine — Gemma-based extraction with an image-preprocessing pipeline and section-by-section validation-retry loops, an earlier iteration of the graph-based system above.
Languages
AI / ML
Document AI / Computer Vision
Infra / Tools
