About
I'm an AI Engineer at Jade & Sterling and an M.S. student in Computer Science and Engineering at North South University. My work sits where agentic AI, retrieval, and applied machine learning meet production constraints — building agent workflows, retrieval pipelines, memory systems, and the evaluation frameworks that keep them honest.
I came to engineering through statistics. Four years of studying estimation, inference, and the ways data misleads you turned out to be unusually good preparation for a field where the hardest question is rarely "does it run" but "how would I know if it were wrong." That instinct shows up in most of what I build: retrieval quality checks, calibrated scoring, deterministic experiments, coverage gates.
Alongside that, I research deep-learning-based visual saliency — evaluating how well saliency models predict fixation behaviour across autistic and typically developing groups. It is a problem where the metrics disagree with each other, which makes it a good teacher.
Strengths
Four claims, and something on this site you can check each one against — a shipped system, a repository, or published research.
- Evaluation-first AI engineering
- I treat the measurement harness as part of the deliverable, not an afterthought. Retrieval-quality scoring, response-consistency checks, and deterministic regression suites ship alongside the systems they judge — because an agent that fails silently is worse than one that fails loudly.
- Agentic systems that survive production
- Retrieval, tool use, orchestration, and persistent memory assembled into versioned services with typed contracts and explicit failure handling, rather than notebooks promoted to prod.
- Statistical rigour applied to ML claims
- A statistics background means I read a benchmark for what it actually measures. My saliency research exists precisely because four standard metrics disagree about which model predicts which population.
- Correctness proofs and engineering hygiene
- My algorithms work ships a brute-force oracle to validate the fast path, tested across four compiler/platform combinations under AddressSanitizer and UndefinedBehaviorSanitizer. The same discipline goes into services.
Technical
“Primary tool” means most days. It is only worth anything because “Familiar” is stamped here too, instead of being quietly left off.
Programming
- Python, SQLPrimary tool
- JavaScript / TypeScript, C++Comfortable
- R, MATLABFamiliar
AI & Agents
- RAG, LangGraph, LangChainPrimary tool
- CrewAI, Vercel AI SDK, Mem0Comfortable
ML & Data
- pandas, NumPy, Evaluation & benchmarkingPrimary tool
- PyTorchComfortable
Backend
- FastAPIPrimary tool
- Django, Next.jsComfortable
Data & Cloud
- PostgreSQLPrimary tool
- MySQL, Google Cloud PlatformComfortable
- DuckDBFamiliar
Tooling
- Git, Linux / UnixPrimary tool
- DockerComfortable
Bangla (Native / bilingual) · English (Full professional) · German (Elementary)
Education
M.S. in Computer Science and Engineering
Graduate coursework chosen around software correctness and experiment design — the two disciplines production AI work keeps asking for.
- Software Quality Assurance: test planning, verification and validation, defect taxonomies, and quality measurement — the formal vocabulary behind the evaluation harnesses I build at work.
- Modeling and Simulation: model design, experiment planning, and validation, including how to distinguish a model that fits from one that generalises.
- Carrying the visual-saliency research thread alongside coursework, treating metric disagreement as the object of study rather than an inconvenience.
B.Sc. (Honours) in Statistics
Four years of estimation, inference, and the ways data misleads you — the foundation I now apply to reading machine-learning benchmarks for what they actually measure.
- Specialised in Data Analysis and Data Mining: data preparation, statistical modelling, pattern discovery, and interpretation.
- Founded the university Programming Club and led it for close to four years — running technical sessions, curating learning paths, and mentoring juniors in programming and Git.
- Served three successive terms in the Career Club, progressing from IT Executive to Head of Information Technology to Head of Administration.
- Built independent projects while studying, including a Markov-chain Discord bot over asynchronous SQLite, and later founded sIndeX.
Higher Secondary School Certificate (HSC), Science
Science group — the mathematics and physics groundwork that later made both statistics and algorithms approachable rather than intimidating.
- Studied mathematics, calculus, physics, chemistry, and biology.
- Competed in science olympiads and science fairs through the college Science Club.
- Began volunteering alongside coursework — a habit that carried through university and into community work.
Certifications
CS50x: Introduction to Computer Science
Certified Python Developer
Introduction to Data Science in Python
The Complete Python 3 Course: Beginner to Advanced
Community
Founding, running, and handing on student technical communities — four years of it, alongside my degree.
Founder
Founded and led a campus programming community — technical sessions, volunteer coordination, curated resources, and mentoring in programming and Git.
Head of Administration
Four years across three successive roles — IT Executive (2019), Head of Information Technology (2020), then Head of Administration (2021–2023) — coordinating administrative operations, event logistics, and cross-team communication while owning the club's digital platforms.
Volunteer
Community initiatives aligned with SDGs 4, 6, 11, and 13, including World Environment Day activities.
Organizing Secretary
Coordinated student programs, communications, and event logistics.
Elsewhere
The quickest way to reach me is email. The code is on GitHub.