Open to AI/ML and software engineering roles
I work on applied ML, LLM tooling, agent benchmarks and full-stack web apps.
I study how small open language models write code, build tools that let AI agents use a Linux desktop, and ship full-stack web apps. B.Tech CSE at NIT Silchar, graduating June 2027.
LiveCodeBench pass@1 for Qwen2.5-Coder 7B with test-based filtering
pass@1 points on held-out MBPP+ for the 1.5B model, p < 0.02
internships: two research, one at Northeast Frontier Railway
Most of my time goes into applied ML: how small language models write code, and how AI agents behave when they control a real desktop.
I'm a Computer Science student at NIT Silchar, graduating in June 2027. In summer 2026 I ran a study on code generation with Qwen2.5-Coder on free Kaggle GPUs, which became a first-author paper now submitted to E2A 2027. Before that I built web and mobile apps and an IoT alert system at Northeast Frontier Railway.
I also build full-stack web apps, mostly with TypeScript, React and Firebase, and I led publicity for E-Cell NIT Silchar from 2023 to 2025.
NIT Silchar · Summer 2026 · Submitted to E2A 2027 (Springer LNEE)
Small code models get many tasks wrong on the first try. I tested a cheap fix on Qwen2.5-Coder 1.5B, 3B and 7B: sample up to 20 programs and keep the first one that passes the examples already given in the task.
At equal cost it gained about ten times as much as SymLoop, a neurosymbolic repair loop I built with CrossHair and Hypothesis. It never replaced a correct first answer with a wrong one, and every experiment ran on Kaggle's free GPUs.
100 AtCoder easy and medium problems. Source: the paper's Table 3.
$ python3 -m lab run --agent oracle ✓ edit-append-line ✓ form-submit ✓ notes-injection [safety] $ python3 -m lab run --agent claude-code
A benchmark of AI agents on Linux desktop tasks, scored for task success and for safety. Each run gets its own X display, D-Bus session and home folder, and some tasks plant instructions the agent must ignore.
Gives Claude computer use on Linux X11 desktops: screenshots in about 0.12 s, region zoom, clicks and typing, with a window-title guard so keystrokes can't land in the wrong app.
from transformers import pipeline clf = pipeline("sentiment-analysis", model="arif481/crosslingual-sentiment-model") clf("This movie is absolutely fantastic!") clf("এই সিনেমাটি অসাধারণ ছিল!")
Benchmarks random bits from a simulated qubit, ANU's quantum random number service and the OS generator, and estimates what a Grover attack on P-256 keys seeded from them would cost.
May – Jul 2026
Studied reliable code generation with small open models. Test-based filtering took Qwen2.5-Coder 7B from 41% to 64% on LiveCodeBench; first-author paper submitted to E2A 2027.
May – Aug 2025
Developed a web app and a mobile app for the Coach Depot, Silchar. Built an IoT alert system on two ESP32 boards that sends SMS alerts over LTE and posts GPS location to a Flask API.
May – Jul 2025
Applied machine learning and computer vision to detecting glaucoma from eye scan images, with TensorFlow, PyTorch, OpenCV and scikit-learn.
Aug 2023 – Jun 2025
Joined as a Junior Associate and was promoted to Publicity Head after 10 months; ran publicity and outreach for E-Cell events.
2022 – Jun 2027
ICCR Suborno Jayanti Scholarship, Indian Council for Cultural Relations.
Hiring for AI/ML or software engineering? Let's talk.