Experience
Work
MIT CSAIL June – August 2026
Undergraduate Research Assistant (UROP), full-time, Cambridge, MA
Researched how to make language models actually follow the explanations they give, with Katie Matton and Mazdak Abulnaga in Prof. John Guttag’s group.
- Built a model-agnostic, test-time wrapper that constrains a model’s answer to stay consistent with its own stated explanation, with no retraining and no access to model internals.
- Tripled causal faithfulness on a hint-following benchmark, from 0.27 to 0.87, beating the prompting baseline; also evaluated on the BBQ social-bias benchmark, across open-weight Qwen and Gemma models.
- Co-first author on the resulting paper, accepted at the Trustworthy AI for Good (AI4GOOD) workshop at NeurIPS 2026.
Software Engineer (UROP), part-time, Cambridge, MA
Worked with Theodore Ehrenborg under Prof. Max Tegmark on formal verification of cryptographic code.
- Wrote machine-checked proofs for Dalek-Lite, a fork of curve25519-dalek, the elliptic-curve library behind Signal and Tor.
- Formally verified 6 scalar and field arithmetic functions, proving each satisfies its specification for every possible input rather than only for tested cases, and proved 34 supporting number-theory lemmas, including Fermat’s Little Theorem.
- Contributed roughly 3,000 lines of proof code across 8 merged pull requests.
Undergraduate Research Assistant, part-time, Cambridge, MA
Applied graph-based models to analyze how cyber risk spreads through industrial control systems and software supply chains.
- Co-author on two papers to appear at the Winter Simulation Conference (WSC) 2026.
Mercor July – August 2025
Mathematics Expert (Olympiad), contract, remote
Solved olympiad-level problems in algebra, geometry, combinatorics, and number theory for datasets used to train and evaluate AI models.
Research Collaborator, internship, remote from Dhaka
Worked on graphlet sampling and link prediction in large networks, contributing to BLANT, an open-source tool that analyzes networks through their graphlets (small connected subgraphs).
- Redesigned two IID sampling methods, Node-Based Expansion (NBE) and Edge-Based Expansion (EBE), addressing two long-standing MCMC limitations: dependent samples and a sequential design that resists parallelization. Derived their overcounting normalization and developed a method to recover absolute graphlet counts.
- Designed a new algorithm for computing MCMC overcounting factors, cutting runtime from 3.2 hours to 5.4 seconds for 8-node graphlets.
- First author on the resulting PLOS ONE submission, and co-author on “Graphlet- and motif-based link prediction in large networks” (Patterns, 2026), where BLANT-Predict beat 13 baselines across 12 networks with up to about a million nodes.
Gonitzoggo August 2023 – July 2025
Software Engineering Intern, remote
Bangladesh’s math and physics olympiad training platform, with more than 30,000 registered users.
Engineering Team Member September 2024 – July 2025
Built and maintained features across the stack (PHP backend on Linux), and redesigned the contest interface and logic for version 3 of the platform.
Developer August 2023 – September 2024
Joined the engineering team after privately reporting a major security vulnerability, then shipped features for contests and user profiles:
- Draft-ownership transfer, so authors can hand off drafts without admin help.
- Contest author credits, from the database migration through to the public contest page.
- In-browser profile picture cropping (Cropper.js).
Teaching and volunteering
Since 2022, most of my volunteering has been training students in Bangladesh for math and informatics olympiads.
Mentor since 2022; Academic Team Member and Problem Setter since December 2024
- Led the mentor team at the camps that trained Bangladesh’s 2025 and 2026 IMO teams, and mentored at the 2026 camp. Those teams won 1 silver and 8 bronze medals.
- Conducted advanced math classes and discussion sessions in algebra, geometry, combinatorics, and number theory, and prepared problem sets and handwritten lecture notes.
- Coached students who later qualified for national teams.
Core Academic Team Member, Mentor, and Problem Setter
- Trained Bangladesh’s team for the European Girls’ Mathematical Olympiad (EGMO) from 2022 and led its camps in 2025 and 2026. From 2022 to 2026 the teams won 2 silver and 7 bronze medals.
- Conducted live sessions and mock tests at the National Girls’ Math Camps.
- Set more than 10 original problems for the National Girls’ Mathematical Olympiad, and have been on the committee that selects the EGMO team since 2022.
NEMOP August 2024 – August 2025
Founder and Mentor
- Founded NEMOP, a year-long online program inspired by India’s Online Math Club, to train top Bangladeshi students for the International Mathematical Olympiad.
- Designed a curriculum with weekly problem sets and an exam every two weeks.
- Taught advanced algebra, combinatorics, geometry, and number theory, and guided students through olympiad strategy; they went on to do well at national and international olympiads.
Thonk & Code September 2024 – March 2025
Co-founder and Mentor
- Co-founded Thonk & Code to help beginner and intermediate students prepare for the Bangladesh Olympiad in Informatics (BdOI).
- Created a curriculum and learning materials for students up to grade 12, and mentored them for BdOI.
Somantoral September 2023 – December 2024
Web Developer and Writer
Wrote articles for a student-run online math magazine for young people in Bangladesh, and built parts of its website. One of my articles is on the Publications page.
Programming contests June 2023 – November 2024
Problem Setter and Tester
Set problems for the Metropolitan University Inter-University Programming Contest (Sylhet Division, 2024), and tested Codeforces Round 878 (Div. 3) and CodeTON Round 7 in 2023.
Gonitzoggo May 2022 – August 2024
Academic Team Member and Tutor
Wrote olympiad-level math problems and taught in online courses and practice programs for students preparing for the national olympiad.
Shamsul Hoque Khan School and College Scout Group January 2017 – January 2018
Scout
Took part in school scout camps and community service events.
Skills
- Programming languages
- Python, C++, C, Rust, JavaScript, Java, PHP, Go
- ML and AI
- PyTorch, Hugging Face Transformers, vLLM, LLM evaluation and benchmarking
- Tools
- Verus, Docker, Git, Linux, GitHub Actions
- Web and data
- Flask, FastAPI, MongoDB, SQLite, pandas, CodeIgniter
- Theory and systems
- Algorithms and data structures, randomized algorithms, graph theory, formal verification, distributed and parallel computing, compilers, MCMC