I'm Pranav Prasanna Venkatesh, a high school senior in Washington who builds things around search and AI. My research on game tree search was accepted as a NeurIPS 2026 workshop poster, and my open-source engine, Cassini, drew 146 of 160 games against Stockfish 19.
Cassini 6.0 on one H100; Stockfish on 1 thread, 256 MB, 5-man Syzygy; 120 s + 1 s.
Move your cursor across the board to see how the current piece moves.
I build AI systems, research tools and medical software.
I've played chess for about ten years, and losing games is what pulled me into search algorithms. That led to MARS, my solo research project, and Cassini, an open-source GPU chess engine. My goal is to build the strongest open-source chess engine.
I also co-founded Thesisly, an AI research workspace with 1,500+ users, and co-created MediConnect, patent-pending software that gives EMTs patients' allergies, conditions and medications. I teach as well: coding at Challenger Coding, Tamil at Redmond Tamil School and chess at Chess4Life.
Achievements
-
NeurIPS 2026
Paper accepted (poster), Workshop on Foundation Models for Temporal Systems. Sole author. International, 2026.
-
Congressional App Challenge
Winner, WA-08, for MediConnect. 1 of ~394 district winners. National, 2025-26.
-
USACO Platinum
USA Computing Olympiad, top ~2% of competitors. The highest of four divisions. National, 2025-26.
-
$100K Azure sponsorship
From Microsoft Azure, backing Thesisly, the AI research workspace I co-founded.
-
1,500+ users
On Thesisly, including UT Dallas professors and a research lab.
-
Patent pending
MediConnect, the emergency medical records software I co-created.
-
$8k+ in prizes
Won building working apps at security, fintech and AI-for-good hackathons.
More awards
| Award | Year | Level |
|---|---|---|
| 1st place, SafeHack | 2026 | Regional |
| Best FinTech Project, Frontier Cascadia | 2026 | Regional |
| Finalist, Open Atlas | 2026 | Regional |
| State qualifier, DECA Entrepreneurship Independent Business Plan | 2025-26 | State |
| State qualifier, TSA Drone Challenge (UAV) | 2025-26 | State |
| College Board National Recognition Program, School Recognition Award | 2025-26 | National |
Cassini
Open-source GPU chess engine
I built Cassini alone and released it publicly under GPL-3.0. It pairs a 188M-parameter transformer evaluator with PUCT search, written in Python and Rust with GPU kernels.
github.com/PranavPrasannaV/cassini-6.0Against Stockfish 19
146 draws and 14 losses in 160 games, about 30 Elo below it in this match.
- Cassini
- Cassini 6.0 on one H100
- Stockfish
- 1 thread, 256 MB, 5-man Syzygy
- Time control
- 120 s + 1 s
Each square is one game: half-filled for the 146 draws, solid for the 14 losses. Sorted, not in the order played.
4th of 25, scoring 20/24.
Chess Engines Diary new-engine exam (Oct 2026).
MARS
Mamba Adversarial Recurrent Search
MARS is a game tree search algorithm I designed and tested on my own, with no mentor or lab. Standard MCTS re-encodes every position from scratch. MARS uses a learned world model to carry memory forward along a line of play.
+52 Elo
Over a PUCT baseline, 95% interval [+34, +70], across 1,040 games.
The axis is the board: each square is 10 Elo.each square is 20 Elo.
Value Magnitude Does Not Survive the Recurrence: Diagnosing and Repairing a Mamba World Model for Latent Search.
The world model's value estimate shrinks toward zero on captures, which I diagnosed and repaired.
The shrinking title illustrates the finding. The chart after it is measured.
Sole author. Accepted (poster), NeurIPS 2026 Workshop on Foundation Models for Temporal Systems, Sydney.
Value is P(win) − P(loss), from −1 to +1, averaged over hanging-piece positions from real games. The first bar comes from an earlier 62-case probe, where the evaluator read −0.425.
The reading comes back too small
MARS plays out lines in a small learned world model instead of calling the full evaluator at every step. When a hanging piece gets captured, the evaluator reads −0.453. The world model reads −0.12. It knows it was punished, not how badly.
Two fixes, both wrong
A cosine of +0.80 with the evaluator's embedding suggested the information was there and the value head was missing it. So I sharpened the value target, twice. Each run moved the reading further from the truth.
The probe that ended it
I decoded the same hidden state through the evaluator's own value head. It returned −0.039, worse again, so a better readout couldn't fix it. A capture emits the same move token as a quiet move, and the scan's decay gate shrinks magnitude at every step.
The repair
The planner already holds the real board, so I compute material from it and train the head on the residual only. It adds no parameters. The reading moved to −0.331 within 3,000 steps and −0.374 in the final model, and rollouts now stay usable across all eight plies the probe measures.
What it's worth
The planner with this model beats a PUCT search on the same evaluator by +52 Elo, 95% interval [+34, +70], over 1,040 games, with less time per move. The repair itself converts to about zero Elo, because the planner's root guard already catches those hangs in play. That guard also removes the positions an on-policy correction needs, by a measured factor of 3.5.
Tested on chess, with one evaluator and a deliberately weak network at 34.4% top-1.
How I ran it
I wrote my decision rules before running matches. I trained on rented spot A100 GPUs and checkpointed every four minutes, because instances could be evicted mid-run.
Thesisly
AI research workspace
I co-founded Thesisly and serve as CTO. It takes students from idea to paper.
1,500+
users, including UT Dallas professors and a research lab.
$100K
Microsoft Azure sponsorship backing it.
MediConnect
Emergency medical records software
I co-created and developed MediConnect, which gives EMTs patients' allergies, conditions and medications from hospital records through Epic FHIR.
Congressional App Challenge winner, WA-08 (1 of ~394 district winners). Patent pending.
- Allergies
- Penicillin
- Conditions
- Type 1 diabetes
- Medications
- Insulin glargine
More experience
-
FoolProof
Hackathon project
An adversarial consumer-protection trainer for financial literacy, with a deterministic grading engine and damage calculations that cite CFPB and FTC sources.
-
Bushel
Hackathon project
A post-wildfire reforestation planner that analyzes burn severity, finds areas beyond natural reseeding range and computes cone requirements by species, seed zone and elevation band.
-
Hackathons
Competitor
Built working apps at security, fintech and AI-for-good hackathons, winning $8k+ in prizes.
-
USACO and Codeforces
Competitive programmer
Reached USACO Platinum, the highest of four divisions.
-
Redmond Tamil School
Lead Developer and Teaching Assistant
Lead the student developers who build and run the school's software. As a TA, manage a weekly class of about 20 and have taught 100+ students.
-
Challenger Coding
Vice President and Instructor
Teach Scratch, Python and Java, and plan lessons and curriculum at a student-run nonprofit whose group has taught 3,000+ students since 2023.
-
Chess4Life
Chess Coach
Teach classes of about 8 kids ages 6-13. I also compete: Chess.com 1770+, top 1% globally.
Skills and tools
| Area | Tools | Where I used them |
|---|---|---|
| Languages | Python, Rust, Java | Cassini, MARS, teaching |
| Machine learning | Mamba state-space models, transformers, distillation | MARS, Cassini |
| Game search | PUCT, MCTS, Gumbel root selection | MARS, Cassini |
| GPU and cloud | Azure, rented A100 and H100 GPUs, spot-instance checkpointing | MARS, Cassini |
| Web and mapping | React, MapLibre, geopandas, rasterio | Bushel |
| Health data | Epic FHIR | MediConnect |