I like building things, learning how they work, and seeing what happens when people use them.
BattleCP
Live product
Asynchronous multiplayer Battleship where solving Codeforces
problems earns you turns. A Rust backend handles the rules and
keeps players in sync over WebSockets.
The server owns game rules and coordinates match state.
Codeforces submission checks connect programming progress to
gameplay, while reconnect handling keeps a match recoverable.
Public launch feedback informs iterations on the rules and
interface.
Bluff
In development
An imperfect-information card game where lying and catching lies
matter. I’m exploring whether agents can learn an opponent’s
habits when they play against the same person over several rounds.
Experiments compare agent variants against scripted opponents
over seeded matches. The current evaluation is synthetic; human
play and broader generalization are not established.
PokeForge
In progress
A Pokémon battle-agent project investigating decision-making with
hidden opponent information, legal action constraints and
reproducible baselines.
Python / PyTorch / Pokémon Showdown
Baseline comparison
Pinned FoulPlay and Metamon baselines complete legal local Gen 9
OU battles. Paired measurement is in progress.
The current focus is reproducible comparison against pinned
third-party baselines. Competitive strength, benchmark-server
conformance and public ladder performance have not been
evaluated.
Drishti-XAI
Research prototype
An explainable diabetic-retinopathy screening prototype combining
retinal image analysis, lesion evidence and Grad-CAM to support
clinician review.
MATLAB / Python / PyTorch / Simulink
What the evaluation showed
Strong internal results did not transfer to an external
population. Domain shift became the central evaluation lesson.
Source private
Evaluation and limitations
The current MATLAB pipeline reports sensitivity 0.908 and
specificity 0.877 on a sealed APTOS test split (733 images).
With its threshold unchanged on Messidor-2, sensitivity falls to
0.363, specificity is 0.865, and AUC is 0.666. These are
separate internal and external evaluations, not clinical
deployment evidence. The repository is private; this summary is
based on the project's current evaluation reports.
Quality gating, vessel/lesion analysis, calibration and Grad-CAM
support the workflow. Clinician assessment and clinical
validation remain open.
To-Do-or-Die
Linux utility
A Rust CLI for todos, with escalating consequences for missed deadlines: notifications, audio and desktop effects. Optional file deletion is disabled by default.
I started LurajBot in eighth grade because I wanted to turn a friend into a Discord bot for our server. It was my first project, and the one that got me hooked on building things with code.
A few projects I’ve worked on with teams through classes, clubs and hackathons.
Algorithima
Openverse
A visual teaching tool I co-built with Openverse contributors, used by professors for DSA and C programming courses. Our team was selected for and presented at OOSC 3.0, IIT Kanpur, in 2025, and presented the tool at T4E 2025, IIT Madras.
An online judge built with a team at IIIT Kottayam. I worked on contest lifecycle, staff workflows, submission dispatch validation and per-test verdict persistence.
A Workspace identity-proof prototype with a Noir circuit and Solidity verifier. Synthetic proof checks pass; the live OAuth-to-local-chain flow is unverified, and the prototype does not establish current student enrollment.
Team Hogorithm's trauma-support prototype, awarded Best Software at IEEE MACE .hack(); '26. I contributed Kotlin safety routing, domain models, UI and tests.
A team image-geolocation project using ResNet, quadtree geographic cells and transfer learning. Experiments compare data efficiency and adaptive cell methods.
I learn best by building things and testing ideas, especially in
algorithms, machine learning and game theory.
Algorithms and systems
I enjoy algorithmic problem solving and digging into what happens underneath: memory, concurrency and the systems details behind the code.
Game theory
I keep coming back to games with hidden information. They’re a fun way to think about how agents reason and adapt when they can’t see everything.
Machine learning and NLP
I’m learning machine learning and deep learning through coursework and projects, with a current interest in how language models behave and how to evaluate them.
AI safety
I’m interested in when a model’s behavior changes, how to notice that change, and how much confidence an evaluation deserves.
Built Edhanta’s learning platform from scratch, including course progression, parent dashboards, authentication and role-based access. Developed the Razorpay payment system end to end, from checkout to subscription activation; the platform reached 20–30 paying users by the end of my internship.
I’m a fan of Star Wars, The Lord of the Rings and Dune, and I’ve read a fair amount of each. I also enjoy games, Pokémon, poker, basketball, trekking and travel.
I like tweaking my Arch setup and do an unreasonable number of typing tests. Monkeytype ↗