Software Development Officer (AI & Automation) at Jamuna Bank PLC, Dhaka Production AI for banking eKYC · Robust / parameter-efficient ML researcher · Fall 2027 PhD applicant in AI/ML
I build systems that have to work in production, and I study why ML models fail when real-world inputs are noisy. At Jamuna Bank I work on digital onboarding, identity verification, payments middleware, and internal platforms. In my independent research I work on robust, parameter-efficient fine-tuning.
Public repos here are sanitized samples — they show design and engineering approach, not production code, models, or customer data.
- PhD, Fall 2027 (US) — trustworthy AI, robust and efficient ML, PEFT, AI safety
- Remote, part-time — AI/ML engineering, model evaluation, computer vision, and full-stack software engineering (~15 hrs/week, evenings Asia/Dhaka)
- Banking / fintech — Software/AI roles building on production eKYC and core-banking integration work
Built a production face liveness / presentation attack detection (PAD) system for digital account opening:
- NID/document-to-live face matching (ArcFace), Swin-B spatial PAD, 3D ResNet-18 temporal liveness, and physics-informed screen/replay detection
- Decoupled PyTorch CUDA inference (FastAPI) from an asyncio gateway — latency cut from over 5s to ~800ms
- Locked-set results: FAR 0.00% (0/33 replay attacks), FRR 3.125% (1/32 live)
- Also: NID OCR, Kafka-based event pipelines for transaction risk flags / AML checks / deduplication, and payment & core-banking middleware
→ Start here: jbl-liveness-verification · banking-engineering-samples
A controlled diagnostic study (LayoutLMv3 + LoRA on FUNSD, matched 5-seed protocol) of PEFT under synthetic document noise:
- GateConcat: +1.51 pp noisy macro-F1 over LoRA under 70% character substitution (p=0.003) and +2.30 pp under 70% OCR confusion (p=0.002)
- Honest limits, reported in the repo: SALLoRA (hand-crafted weighting/masking) did not beat LoRA; a constant-gate capacity control captures part of the GateConcat gain (p=0.21 vs that control); the learned gate does not detect corruption (AUC ~0.52); gains do not transfer to layout corruption
- I value this kind of result — capacity-matched controls and negative results included, not just headline gains
Research interests: parameter-efficient fine-tuning, robust model adaptation, trustworthy / responsible AI, uncertainty quantification, federated learning, document understanding, NLP, computer vision, multimodal robustness under structured corruption
Publications: 4 co-authored papers at IEEE ICCIT 2023 — suicidal thought detection with NLP, privacy-preserving federated learning for lung cancer classification, PCOS detection with ML, and YOLOv8 object detection for self-driving cars.
| Repo | What it shows |
|---|---|
| gateconcat-peft | Research code + full experiment artifacts for the PEFT robustness study above |
| jbl-liveness-verification | Liveness / PAD design for banking onboarding (sanitized) |
| jbl-face-match | NID-to-live face matching design (sanitized) |
| banking-engineering-samples | Index of sanitized Jamuna Bank systems: middleware, eKYC ops, payments, internal tools |
| Personal-website | My portfolio site — rafiimam.github.io/Personal-website |
| agent-capture-bench | Benchmark for AI-agent recorder capture fidelity vs independent ground truth |
- Languages: Python, TypeScript, JavaScript, C/C++, SQL
- AI/ML: PyTorch, TensorFlow, CUDA, ArcFace, LayoutLMv3, LoRA/PEFT, OpenCV
- Backend: FastAPI, Node.js, NestJS, ASP.NET Core, REST, WebSockets, asyncio
- Frontend / mobile: React, React Native, Next.js
- Data & infra: Apache Kafka, PostgreSQL (pgvector), MongoDB, SQL Server, Docker, Kubernetes, Linux, Firebase
- B.Sc. in Electrical & Computer Engineering, RUET (2023) — RUET Merit Scholarship 2020, 2021, 2022
- IELTS Academic 7.0 · GRE 304 (Quant 164)
- At Jamuna Bank PLC since December 2023
- Email: rafiimam96@gmail.com
- Website: rafiimam.github.io/Personal-website
- Location: Dhaka, Bangladesh (Asia/Dhaka) — open to remote work