Software Engineer with 4+ years of experience in machine learning and backend systems. At Dutch-Bangla Bank PLC I lead R&D on LLMs, RAG and document AI, and ship production systems for fraud detection, credit assessment and cross-border payments. Outside work I build full-stack AI apps, Flutter apps and developer tools.
Currently
Building RAG and document-AI pipelines for banking
LLM-based OCR extracts data from CIB reports, bank statements, loan applications and NID cards with 95%+ accuracy.
CIB extraction and credit scoring for automated loan approval reduced processing time by 75%.
Cross-Border Transaction Automation
SWIFT MT103, MT940, MT950, MT942 and MT910 parsing, matching and rule-based classification, reconciling statements and credit confirmations against payment instructions. Cut reconciliation from hours to minutes and removed 90%+ of manual effort.
Led migration from legacy MT103 to ISO 20022 pacs.008 for cross-border payment compliance.
Built and maintained scalable REST APIs with Python and Django REST Framework.
Created an NLP-based product mapping model for product comparison and recommendation.
Built and published cross-platform Flutter apps for internal business operations.
PythonDjangoDRFNLPFlutter
Projects
Things I've built
Side projects and freelance work: AI agents, document AI, full-stack apps and Flutter apps. Production systems from my day job are under Experience.
Showing 6 projects
AI / MLIn progress2026
Photon
Local-first desktop AI teammate. An Electron app runs sandboxed tools (files, shell) inside your workspace while a Django server handles auth, encrypted secrets and routing across LLM providers, with an approval step before any write or shell command.
Open-source, self-hosted session replay and heatmaps for Django. Text is masked in the browser and no IP addresses are stored; 15 documented rules flag rage clicks, dead clicks and form abandons for review in Django admin.
Turns the password-protected statement PDF that bKash emails into a spending dashboard: balance over time, cash flow, fees, top recipients and when you transact. The PDF is decrypted and parsed entirely in the browser; nothing is uploaded or stored.
Extracts transactions from PDF bank statements of any layout using vision LLMs. Pages are processed in parallel, validated with Pydantic and returned with confidence scores and token-cost tracking.
Python library that turns a pandas DataFrame into a quality report: completeness, types, statistics, correlations and plots, exported as a Word document.
Resume builder app to create, edit and preview resumes and export them as PDF, in English and Bengali, backed by a Django REST API shared with JobBoard. My most-starred repository, with 18 stars and 8 forks.