About
I'm a senior software engineer at KONA Software Lab in Dhaka, Bangladesh. I've been an engineer for six years, most of them in banking and payments: card platforms for banks, a token-backed commerce platform on Hyperledger Fabric, and now the AI assistants that run on top of them.
Right now I work on the agentic AI layer of KONA's card management system. I wrote most of the MCP tool servers the agents work through, the approval flow for writes that had no human check of their own, and the eval suite we test it with. We then took the same stack into the Nagad app as KON-AI, which is integrated and in testing. The models run on our own hardware, so bank data never goes to an outside model API.
Before that I was the top contributor to NCP, a Hyperledger Fabric commerce platform that runs for KONA I in Korea, and I worked directly with the Korea team, in English, on requirements, demos and releases. Earlier I did Hyperledger Fabric R&D at BJIT and built React and Node.js web apps.
I work mostly in Java and Spring Boot, TypeScript and Node.js, and Python, with some Go. I use Claude Code every day and build small tools around it. I studied Computer Science and Engineering at Daffodil International University, where I also spent a lot of time on competitive programming.
How I work
Safety rules are enforced in code, not left to the prompt.
A prompt can be talked around. KON-AI's PIN rules are enforced by the screen finder itself, so a forgotten PIN never leads to the change-PIN screen.
Where it shows →The bank and the user come from the request, not from the model.
An agent reads them from the request headers, so a prompt injection cannot pick a different bank.
Where it shows →The agent drafts a change and a person approves it.
Writes wait for a person, and an approval is a single-use token bound to the exact change. You can try it in the Lab.
Where it shows →I measure with evals before tuning.
An eval suite of 2,265 cases shows what broke, and I trace each failure back to its root cause.
Where it shows →New data is merged in, so a fetch cannot remove what was there.
In whichllm, live data is laid over the curated snapshot instead of replacing it.
Where it shows →The expensive model is used only where the cheap one is unsure.
NID KYC reads every card with OCR and asks a stronger model to re-read only the fields the validators doubt.
Where it shows →
Path
2025 – now
Senior Software Engineer, KONA Software Lab
The agentic AI layer of a card management system and the Nagad app, the self-hosted LLM stack under it, and fund tracing for KONA I.
2022 – 2024
Software Engineer, Level 2, KONA Software Lab
Top contributor to NCP, a token-backed commerce platform on Hyperledger Fabric, that runs for KONA I in Korea.
2021 – 2022
Software Engineer, BJIT
Hyperledger Fabric R&D in Go and Spring Boot, and an Ethereum courier dApp.
2020 – 2021
Web Developer, iTech Soft Solutions & Semicolon IT Solutions
React and Node.js web apps on PostgreSQL and MongoDB.
2017 – 2021
B.Sc. in CSE, Daffodil International University
Competitive programming on Codeforces, ICPC Dhaka Regional 2018, and 1st place at the DIU programming contest in 2019.
What I work with
- Languages
- Java, TypeScript/JavaScript, Python, Go, SQL, Solidity, C++
- Backend
- Spring Boot (WebFlux, Batch, Authorization Server), Node.js/Express, FastAPI, Microservices, RabbitMQ, Socket.IO, Kafka Connect
- AI and LLMs
- Agentic AI, LLMs, RAG, MCP (Java and TypeScript SDKs), LangGraph, vLLM, LiteLLM, Qwen, RAGFlow, Embeddings and rerankers, Gemini, Google Cloud Vision, Evals and guardrails
- Frontend
- React, Angular, Next.js, Svelte
- Mobile
- Kotlin Multiplatform, Compose Multiplatform, Swift, Capacitor
- Data
- Oracle (star schema), PostgreSQL, Redis, MongoDB, MinIO, Spring Batch, Apache Superset, JasperReports
- Security and blockchain
- OAuth2/OIDC, WebAuthn passkeys, Okta, HSM (PKCS#11), X.509, Hyperledger Fabric
- DevOps and testing
- Docker, AWS EC2, Jenkins, PM2, Gradle, Maven, JUnit, WireMock, pytest, Jest, Playwright, TestNG, RestAssured, Selenium, Hyperledger Caliper
- AI coding tools
- Claude Code (daily), earlier Google Antigravity and GitHub Copilot
Competitive programming
- Codeforces: 286 rated contests (2017–2025), 800+ problems solved, max rating 1414 (Specialist), hardest problem solved rated 1800.
- ACM ICPC Dhaka Regional 2018: 96th place.
- DIU Intra-University Programming Contest 2019: 1st place.
- Executive, DIU Computer & Programming Club (2017–2019).