Hi, my name is
Arbab Chowdhury
AI & Data Engineer for regulated financial systems.
I build regulatory-grade data platforms and production RAG / GenAI systems for banks and compliance-heavy teams — with the correctness, auditability, and rigor that regulated data demands.
Résumé See my work Get in touch01. About
Background
I'm a senior data engineer and software architect at PNC, a top-10 U.S. bank, where I make sure financial and regulatory data is correct, compliant, and on time. I designed a relational data model that has run in production without a failure, and I validate Basel III / LCR liquidity calculations against source-of-truth reconciliations in OFSAA.
Now I build the AI layer on that same rigor: retrieval-augmented generation over regulated documents, AWS Bedrock pipelines, and agentic workflows — with the citation discipline, evaluation harnesses, and audit trails that regulated data requires. As a CISA, I design those systems to be audited, not just to work.
I bring enterprise depth to AI work and AI leverage to enterprise work, and I make pragmatic build-vs-buy calls across AWS, Azure, and GCP.
02. Work
Selected projects
BankOS — Interactive Model of a Modern Bank flagship
An explorable, explainable model of a Tier-1 bank: legal entities, lines of business, accounting, treasury and liquidity, risk, capital, controls and regulatory reporting, connected by a transaction-to-report digital thread. Built on a provenance-tracked reference corpus of public regulatory filings (Call Reports, FR Y-9C, SEC filings), a canonical chart of accounts with a source-cited crosswalk to public reporting lines, and a graph-based visual explorer.
RegIntel — Regulatory Intelligence Platform flagship
Citation-grounded RAG assistant over banking regulation (Basel III/LCR, FFIEC, OCC). AWS Bedrock-native — Titan embeddings, Claude Converse, hybrid retrieval, refusal-aware eval harness.
RAG, three ways flagship
One document-Q&A use case shipped three ways — no-code (n8n + Supabase), managed cloud (Bedrock Knowledge Base), and custom (Python RAG) — with an honest cost / latency / control trade-off table. How I make build-vs-buy calls, in code.
Agentic operations platform
A production multi-agent system I run daily: scheduled Claude agents triage email and reminders and draft work for human review, a persistent knowledge graph (via MCP) gives every agent shared memory, and a zero-LLM operations dashboard on Cloudflare Pages Functions + D1 keeps the deterministic parts cheap and auditable.
HIPAA web-compliance audit — multi-site healthcare practice
Hands-on privacy and security review of three patient-facing websites: live testing of form-submission pipelines and third-party tracker behavior, a cross-site gap matrix, and prioritized punch lists the developers could act on directly.
AI booking assistant — telehealth medical group
Conversational AI front desk for a two-office primary-care practice: office-aware routing, appointment booking, SMS compliance (A2P 10DLC) and an audit of the inherited automation workflows.
ComplianceIQ
The same retrieval architecture applied to a second regulated domain. Hybrid (dense + BM25) retrieval with reciprocal-rank fusion, citation enforcement, and Claude-as-judge faithfulness evaluation.
More on github.com/arbabc-ai. Client work is described without client names.
03. Credentials
Certifications
CISA — Certified Information Systems Auditor
ISACA
Artificial Intelligence: Implications for Business Strategy
MIT Sloan School of Management & MIT CSAIL (Executive Education)
04. Writing
Notes & teardowns
Coming soon — architecture notes, build-vs-buy decisions, and RAG-evaluation writeups from regulated-data work.