Sai Maruvada

Autonomous nanodrone research

ML lead · Advisor: Prof. Aloysius Mok, UT CS · Aug 2026–present

Person detection for a nanodrone that follows people. Trained 8-bit quantized CNNs on 118K images to fit a 64 mW processor, scoring an F1 of 0.801 against 0.789 for the prior released model, both quantized. Retrained on targeted negatives to cut false alarms on pets and mannequins from 24% to 8%. Fixed a silent code-generator bug that zeroed negative chip weights and added 5 release checks. Ported the Crazyflie simulator to macOS and built a closed-loop person follower that passed 5 safety tests.

P.R.I.S.M. 3D

Co-author · ML engineering

3D deep-learning pipeline predicting lung cancer survival from CT scans, co-developed as the second of three authors. 0.75 AUC across 422 scans, 0.946 Dice on nnU-Net segmentation. Validated on 58 external cases with Shanghai Pulmonary Hospital. Won the Citadel Securities Innovation Prize.

The Honest Backtester

Solo · quantitative research

A backtester that argues back. My own momentum strategy looked like a winner until I audited it, so I built the audit into the engine where it cannot be switched off. Six checks run on every result: selection bias re-tested on the full S&P 500, a cost sweep scaled to measured turnover, a 30% holdout, enforced lookahead prevention, regime coverage, and runs reproducible under a content hash. Backtested over 40 months, my momentum rotation posted a 136% CAGR, but on a watchlist I had picked with hindsight, with a 45% drawdown against SPY's 8%. The big return was mostly selection bias.

Zone

Co-founder & CFO

iOS app that blocks distracting apps by location, using BLE and GPS geofencing, built in SwiftUI. Led the five-person team that shipped the prototype. Won the Congressional App Challenge and TYE Regionals, placed 7th at TiE Global, and was a Diamond Challenge finalist. Secured $4,000 in funding.

SaiSneakResale

Founder · sole proprietor

Sneaker resale business started in June 2022 and run solo until March 2026. Automated purchasing for limited releases and priced from comparable-sales data. $35,000 revenue in the first year and $45,000 in total.

Prediction market data

Solo · market infrastructure

Always-on collectors that poll Kalshi's API about every 42 seconds for order books, with a second feed for sports contracts, running on a Linux server. From July 30 to August 17, 2026 they logged 8.06M order-book snapshots across 54,380 crypto contracts. I worked the fee math before the strategy and found Kalshi's fees ate the entire edge of the momentum signal I was testing, so I killed it and built the data asset instead. Future signals get tested against real depth and spreads rather than mid-price assumptions.