Jan 2025 – Present
Senior Software Engineer
KONA Software Lab Ltd. · Dhaka, Bangladesh
I work on the agentic AI layer of KONA's card management system, the assistants built on it, and the self-hosted models underneath.
- Bank operators ask the card management system about products, cards and transactions in English or Bangla, and the assistant finds the right screen or prepares the change. Nothing it prepares takes effect until a person approves it. Read more →
- Designed and wrote most of the MCP tool servers the agents work through: 31 Java/Spring Boot tools in three tiers plus an action catalog for rarer operations, a Node.js screen finder over the portal's 156 screens, and around 700 tests. Read more →
- Writes need a person's approval. The agent can only save drafts or act as the maker in the bank's maker-checker flow. Writes with no approval step of their own wait for a person in an approval ledger I built on Spring WebFlux and PostgreSQL, with single-use tokens bound to the exact change. Read more →
- Added guardrails to the Python orchestration service (FastAPI, LangGraph): PII redaction on the output stream, input screening, limits on tool calls, and a check for numbers the model states without looking them up. Read more →
- Wrote the eval suite for the whole stack: 57 suites and 2,265 cases covering guard checks, tool calls and chat. The latest run passed 98.7% of guard checks. Read more →
- Took the same stack into the Nagad app as KON-AI, now integrated and in testing. Split prompts, configuration and builds into a shared base plus per-product profiles, wrote the screen finder that matches Bangla, English and Banglish questions to app screens with PIN-safety rules in code, and added the screen-card event that drives a button in the app. A judge-backed guard cut wrongly blocked questions from 10 of 146 to none. Read more →
- Set up our self-hosted LLM stack on an NVIDIA DGX Spark: vLLM serving a 27B Qwen model, embeddings and a reranker behind a LiteLLM gateway. Load-tested it at about 205 tokens/s across 16 streams and moved the agent services onto it. Read more →
- Built the RAG system our card personalization engineers use as their daily reference: a customised RAGFlow over 54 card specifications (EMV, GlobalPlatform, Mastercard M/Chip, Visa). Read more →
- Built a dashboard API for card issuers on our Oracle star-schema warehouse, made all 15 cardholder reports tenant-aware, and hardened the Spring Batch ETL with per-table load tracking and rerun safety. Read more →
- Built KONA Tag Explorer, backend and frontend, with one teammate: fund tracing for a KRW-pegged card system where every transaction becomes a UTXO with a lineage tag. I wrote the React explorer, most of its APIs, payment-depth contribution scoring and month-end UTXO consolidation. In production. Read more →
- Co-built KONA's identity provider on Spring Authorization Server, with passkey (WebAuthn) login and Okta federation, and wrote its React passkey client. Read more →
- Built tooling to run and deploy the stack: a one-command local runner for the services, an auth test harness, and deploy jobs that fail if the old version is still serving. Read more →
- Wrote design docs for the stack, including the high-level design for the agentic AI platform and the product-profiles design, planned and owned the agentic-CMS sprint backlog, and reviewed and merged the team's changes to the stack.