Open to research and AI evaluation work

MahdiHasan

Independent AI Evaluation Engineer · Dhaka

I research how to detect deception in language models, and design the benchmark tasks that train and test AI coding agents.

Mahdi Hasan

AI evaluation in numbers

280+

approved benchmark tasks for AI coding agentsWritten for AfterQuery Experts and Snorkel AI, under NDA

  • 130+

    SWE-bench-styleReal bugs and features from my own TypeScript, JavaScript and Python repos

  • 150+

    Terminal-BenchLinux tasks in ML, compilers and parsers, and Terminal-Bench 2.0

Papers and preprints

Papers and preprints, with the full index keeping every abstract and link in one place.

Ideas in practice

Featured builds, with the full story behind each one on its own page.

An arc diagram of one song — one-second segments along a timeline, joined by arcs wherever two moments sound alike

A graph neural network reads a song's repetition structure, BERT reads its description, and cross-attention lets each side look at the other.

  • PyTorch
  • PyTorch Geometric
  • BERT
  • GATv2
  • librosa
The UniCommunity sign-in screen beside an illustration of the BRAC University campus

A campus platform for faculty reviews, course resources, discussions, and finding project teammates.

  • React.js
  • Node.js
  • ExpressJS
  • MongoDB
  • RESTful API
  • Google OAuth 2.0
A graphic reading Get Started Contributing to GitHub with GitHub and branching icons over a dark network pattern

Small Python projects designed to make a first open-source contribution less intimidating.

  • Python
  • GitHub Actions
  • Open Source

Along the way

Selected chapters from the longer professional and academic timeline.

2025 — Present

Contract Engineer, AI Evaluation

AfterQuery Experts · Remote

Wrote 130+ approved SWE-bench-style tasks from real bugs in my own repositories and 120+ approved Terminal-Bench tasks in ML and compilers. Each one ships fail-to-pass tests, a reference fix and a pinned Docker image, and has to hold up against frontier coding agents.

2025 — 2026

Benchmark Task Author

Snorkel AI · Remote

Wrote 30+ Terminal-Bench 2.0 tasks: Docker and Linux environments with realistic failures and automated grading. Graded a frontier lab's code-review model on real GitHub pull requests and wrote rationales used as RLHF data.

2024 — 2025

Training Specialist

Outlier · Remote

RLHF on front-end coding across several projects: ranked, rewrote and critiqued model responses. Audited a frontier assistant's extension calls (maps, notes and more) for the right call and a grounded answer.

2022 — 2026

BSc in Computer Science and Engineering

BRAC University

Graduated in 2026. Thesis on catching lies in a 70B language model from its internal activations. Coursework across AI, systems, electrical engineering and math.

Notes on code and life

Featured writings, experiments, and occasional detours.

উদ্দেশ্যহীন

এই তো জীবন কুট্টুস, শান্ত কয়েকটা ছুটির দিন।

5 min read

কী আছে জীবনে?

টাকা-পয়সা, গাড়ি-বাড়ি আর সব রিসোর্স থাকার পরও যদি জীবনের ছোট ছোট মুহূর্তগুলা ট্রুলি এনজয় করার সময় না থাকে, তাইলে আসলে কী থাকে?

10 min read

Your First Open-Source Mystery

Know Python basics but stuck on what to build next? This blog turns that ‘blank screen’ moment into a simple first open-source win—where even others can help improve your code.

15 min read

The Weight of Soft Things

There are no right or wrong decisions in life—only experiences that shape us and make us stronger.

2 min read

Off the clock

Video and photography. A warmer counterpoint to the research half of the site.

Start a conversation

Have a project, research idea, or collaboration in mind? Send me a note and I’ll get back to you soon.

Location
Dhaka, Bangladesh