Evaluation is the real skill.
Anyone can call an API. The hard part is knowing whether it works — golden datasets, hallucination detection, LLM-as-judge, and drift monitoring are what separate prototypes from systems you can trust in production.
Available for AI/ML roles
Kaggle Master · #29 global
First-author research
Shipped production ML
Bengali speech AI
Build the future.Ship what matters.
I design and engineer production AI systems with measurable business impact.
AI/ML Engineer specializing in Bengali Speech AI, production ML, and GenAI platforms. First-author researcher. Kaggle Top 1%.
I turn ambitious ideas into products teams can trust in the real world.
Design Principle
Build → Deploy → Measure → Improve. Great AI systems are reliable, observable, and cost-effective under real constraints.
I optimize for what matters in production: evaluation quality, inference latency, and system trust.
Philosophy
Hard truths about building AI that ships.
Production thinking beats accuracy.
A model that scores 0.95 in a notebook but can't handle latency, cost, or scale is a liability. I optimize for the full stack: serving, observability, failure modes, and the infrastructure that keeps AI reliable under real load.
Ship, don't slide.
95% of AI roles are applied, not research. I focus on end-to-end delivery — from data pipeline to deployed API — because companies hire engineers who can build, deploy, scale, and measure. Not just present.
Project Selection
Selected Work
A curated collection of systems I've built. These projects are selected to demonstrate end-to-end engineering—from custom deep learning architectures (Transformers from scratch) and agentic AI middleware, to production pipelines that solve real business constraints like latency, scalability, and vendor lock-in.
Capabilities
Production ML Systems
MLOps & Cloud
GenAI & LLM Systems
Languages & Infrastructure
Data & Search
Domain Specialization
Profile
I'm an AI/ML Engineer from Dhaka, Bangladesh. My journey started during Computer Science studies at BNIST, where courses in linear algebra, probability, and data structures sparked a deep curiosity about how machines learn from data. That curiosity turned into 2+ years of building production ML systems — from end-to-end pipelines with ZenML and MLflow to Dockerized inference services with FastAPI.
I earned a Kaggle Master rank across 22 competitions, published a first-author conference paper on Bengali speaker diarization (BUET CSE Fest 2026), and founded Toolly — a community-driven AI tool discovery platform. My recommender system work delivered a 10% sales lift for a client in 3 months.
When I'm not training models, I explore the frontier of Generative AI — LLMs, RAG pipelines, and LangChain/LangGraph agents. I believe the best ML work happens at the intersection of strong engineering and genuine curiosity.
Education
B.Sc. in Computer Science & Engineering
BNIST
CAREER
The arc.
2026
First-Author Research Paper Published
publicationBUET CSE Fest 2026
Bangla Diarizz: Domain-Adapted Bengali Speaker Diarization via Knowledge Distillation. DER 0.19 · 56% inference speedup · 3.4× real-time on CPU. Read paper →
2025
Kaggle Master
achievementKaggle
Global rank #29 · Top 1% out of 4,082 competitors. 22 competitions total, including Top 1% in Road Accident Risk and Top 2% in BPM Prediction.
2024
Deployed Production Recommendation System
projectFreelance / Client Project
Built and delivered a hybrid recommendation system (collaborative + content-based). Successfully deployed and achieved a +10% sales increase in 3 months for the retail client.
2024
Launched Toolly
projectPersonal Project
Built and launched an AI tool discovery platform (toolly.tech). Defined product vision, led full-stack development, implemented submission moderation and analytics. 400+ curated tools across 15 categories.
Research
Published work
Domain-adapted Bengali speaker diarization via knowledge distillation
A lightweight pipeline for Bengali long-form audio that reaches DER 0.19, with a distilled student model that runs at 3.4× real-time on CPU and roughly 56% faster inference than the baseline — aimed at deployments without heavy GPU infrastructure.
Read paperRight now
What I’m focused on
Building
RAG evaluation pipelines with LLM-as-judge scoring, golden datasets, and hallucination detection — the measurement layer most AI systems skip.
Deploying
Containerized ML services on AWS with FastAPI, Docker, and CI/CD — production inference that handles real traffic, not just localhost demos.
Exploring
Agentic workflows with LangGraph — multi-step orchestration, tool calling, and the evaluation challenges that come with autonomous AI systems.
Available
Open to AI/ML engineer roles, production ML consulting, and research collaborations where the bar is shipping, not slides.
Kaggle Master — global rank #29, top 1% across thousands of competitors. View Kaggle profile
CONTACT
Let's make
something
that matters.
I'm open to AI/ML engineer roles, production ML consulting, and research collaborations. If you want a system that actually ships — reach out.