RKRitwika Kancharla Email

Ritwika Kancharla

Hi, I’m Ritwika. I’m an ML engineer.

I like the whole LLM stack: training models, making them fast, and getting them to solve hard problems.

Right now I’m getting an open 27B model to play games it has never seen and to solve grid puzzles by writing programs, for the ARC Prize 2026. I’m also competing to train the best language model I can in 30 minutes on an AWS Trainium chip. You can read about each one in detail below.

View projectsAbout meGitHub
A painted Siamese cat holding a glass of iced coffee against a deep pink background with pressed flowers.

Projects

Each project page states the problem, describes the approach, and reports results together with their evaluation setting and limitations.

The game ft09 being played by the model, level by level.The game sb26 being played by the model, level by level.
2026OngoingARC Prize 2026

Verified World Models for Interactive Reasoning in ARC-AGI-3

ARC-AGI-3 places an agent in a small game with no instructions. An open 27B model proposes the rules of the game, and a deterministic harness handles perception, records every transition, tests each proposed rule against that record, and executes plans under guard conditions.

6 public games complete (41 of 41 levels), without fine-tuning. Development-set result; not a leaderboard score.
An ARC-AGI-2 test input grid divided into regions.The expected output grid, with each region filled by nested coloured rings.
2026OngoingARC Prize 2026

Verified Program Induction for ARC-AGI-2

A second solver for abstract grid puzzles. The model writes each puzzle's rule as a program, and the program must reproduce every training pair exactly before its answer is considered. The study measures how often such verified programs are still wrong on the test grid.

33.89 leaderboard score (transductive baseline). Verified programs are correct on 10 of 15 test grids.
2026Ongoing

Trainium of Thought: Model and Kernel Co-Design under a 30-Minute Training Budget

My entry in the AWS Trainium Frontier Competition on Devpost, a language-model training competition on a single Trainium2 chip in which wall-clock time is the only budget. The architecture and the fused kernels were developed together so that more useful optimizer steps fit into 30 minutes.

1.3951 to 0.9656 val bpb
Phase 1 ongoing. Currently 2nd on the public leaderboard; 1st at its peak.
2026Ongoing

Imitation Learning for a Long-Horizon Economic Game

Recurrent behaviour cloning on a 719-decision farming economy. A strict evaluation audit showed that the first results were inflated by games in which the agent had failed, and it identified the cause.

100,660 expert decisions
140 complete games with seed-disjoint splits.
A woodblock-style print of a tabby cat asleep on a deep blue background.

Contact

For questions about any of these projects, or to discuss research and engineering work, send me an email.

Email meGitHub