Work
Eleven projects, most of them from KONA Software Lab.
At KONA
AI assistant for the card management system
Bank operators ask about products, cards and transactions in English or Bangla. The assistant finds the right screen or prepares the change, and nothing takes effect until a person approves it.
MCP tool servers for the card system
KONA's agents reach the card system only through MCP tool servers, and I designed and wrote most of them: 31 tools in three tiers, an action catalog for less common jobs, and a screen finder over 156 portal screens. The same servers are used for Nagad.
Guardrails and evals for the agents
Input screening, PII redaction, limits on tool calls, and a check for numbers the model states without looking them up. I also wrote the eval suite that tests them: 57 suites and 2,265 cases.
KON-AI: assistant in the Nagad app
The same agent stack inside the Nagad mobile wallet, for customers and guests. It answers in Bangla, English or Banglish and points people to the right screen. The PIN-safety rules are written in code.
Self-hosted LLMs and a card-spec RAG
I set up KONA's LLM stack on an NVIDIA DGX Spark, put the models behind one gateway and moved our agents onto it. I also built a RAG over 54 card specifications that our card personalization engineers use every day.
Tag Explorer: fund tracing for a card system
Fund tracing for a KRW-pegged card system. Each transaction becomes a UTXO with a lineage tag, so auditors and regulators can follow a balance back to its source.
Reporting and data for the card platform
A dashboard API for card issuers, cardholder reports scoped to each bank, an ETL that records what each load did, and Superset dashboards built from code, on KONA's Oracle star-schema warehouse.
NCP: commerce platform on Hyperledger Fabric
A token-backed commerce platform for KONA I. I worked on offline transaction signing, HSM-backed identities and messaging that recovers after a broker outage, and was its top contributor.
Personal and open source
NID card OCR for KYC
Reads KYC fields from Bangladeshi NID cards in Bangla and English. Cloud Vision reads the text, a small Gemini model parses it, and a larger one re-reads only the fields the rule checks flag, from crops of the photo.
Hisaab: personal finance app
Reads bKash, Nagad and bank SMS, encrypts the data end to end, and runs its AI on the phone.
whichllm: benchmark scraper fix
whichllm ranks local LLMs partly by a benchmark it scrapes. The site changed format, every fetch failed without an error, and rankings ran on an old snapshot. I parsed the new format and made sure a live fetch can only add models.