My Projects

Projects I have worked on recently.

Chasing myelin — how an autism hypothesis about supply became one about timing

Myelin is the fatty insulation around axons, and the prefrontal cortex is the last part of the brain to get it. That makes a tempting hypothesis: if the machinery for making it is constitutionally weak, the last place to be insulated should be the first place to fail — which happens to be the region whose functions look most like the ones disrupted in autism. This is the log of testing that idea on public data, the two reframings the data forced, and the sharper question that came out the other side.

Bendy bodies, restless minds — an ordered connective-tissue lattice and a neural network, joined by a dashed link marked with a question mark, above the measured associations and the explanations ruled out

Bendy bodies, restless minds — a correlation we can measure but cannot yet explain

People with unusually flexible joints get migraines about twice as often as everyone else, and are diagnosed with ADHD around five times as often. The association is real, replicated, and partly genetic. This is the log of an attempt to find the mechanism behind it — the elegant explanation that turned out to be wrong, the things we could rule out, and the one thread still worth pulling.

Where is the blueprint in the C. elegans connectome — a decomposition of reproducible wiring into contact geometry, developmental timing, and a small transmitter bias, with the falsified hidden-code hypotheses

Where is the 'blueprint' in the C. elegans connectome?

A worm's genome is far too small to store its ~7,000 wiring connections one by one — yet the wiring is nearly the same in every animal. So where is the information that specifies it? This is the log of a systematic hunt: a dozen candidate 'hidden codes' tested and ruled out, and the surprisingly simple thing that was left standing.

C. elegans Tier4 substrate anatomy — 3D rendering of the whole-cell electrical model

Tier4 Substrate Anatomy — a whole-cell electrical substrate, rendered from code

An interactive 3D anatomy of the C. elegans Tier4 whole-cell electrical substrate. Every channel, receptor, pump, transporter, gap junction, and ion compartment actually present in the Brian2 model is rendered 1:1 from code with an honest ON / default-OFF / ORPHANED / NOT-INTEGRATED status.

Canonical-AMR top-3 across 3 days: 0.17, 0.00, 0.403

Antibiotic Resistance — Mutation Trajectory Prediction

Predicting the next resistance mutation given an ordered evolution history, over ALEdb plus CARD plus deep mutational scanning data. A LightGBM + autoregressive transformer + reranker stack lifts canonical-AMR top-3 from 0.17 to 0.403 across 14 drug classes.

Cross-cohort AUROC: 0.217 Day 2 versus 0.850 Day 3 on GSE122126

cfDNA Cancer Detection — A Cross-Cohort Recovery Story

Cell-free DNA methylation cancer detection over the Loyfer atlas. Cross-cohort AUROC on GSE122126 went from 0.217 (below random, model inverted) to 0.850 across four of five cancer types. The four mechanisms that closed the gap, and the literature check that showed they're not novel.

P3/time AUROC across splits: Day 1 baseline vs Day 3 leak-clean fusion

Clinical Trial Failure Prediction

Phase-3 trial outcome prediction over the AACT clinical trials database, with a leakage audit that proved the headline number is signal, not label encoding. Audited AUROC 0.8603. The audit methodology is the part worth carrying forward, not the model.

Halothane vs hexafluoroethane network-state response on the C. elegans Cook 2019 connectome — anesthetic pushes command-interneuron firing into quiescent regime, non-immobilizer does not, across four orders of magnitude in dose.

Anesthesia Pipeline — Network-State Validation

Network-state validator for volatile anesthesia on a substrate-agnostic LIF integrator. Predicts EC50s, mutant magnitudes, and Eger non-immobilizer specificity from literature-grounded multi-target perturbation profiles across three independent organism calibrations (worm Cook 2019, fly Winding 2023, mouse generic random graph).

ARC-AGI puzzle showing a blue input grid transforming into a green output grid, with a multi-step transformation strip — Rohit Ravi's ARC Prize 2025 entry

ARC Prize 2025

Ongoing attempt at the ARC-AGI benchmark — a pipeline-based approach to program synthesis, symmetry reasoning, and abstraction discovery.

Illustrated C. elegans worm with colored neural-network overlay — Rohit Ravi's connectomics + transcriptomics project

C. elegans Connectomics + Transcriptomics

Multi-modal analysis asking how much of local wiring in a small nervous system is determined by topology vs. gene expression.

Polymer chain mapped to a scatter plot of molecular features vs. predicted property — Rohit Ravi's NeurIPS Polymer Prediction Kaggle entry

NeurIPS Polymer Prediction

Kaggle competition for predicting polymer properties from SMILES-style representations — a multi-phase hybrid pipeline combining gradient-boosted and deep models.

Side-by-side illustration of ground-truth and predicted RNA 3D tertiary structures with a base-pair / hydrogen-bond legend — Rohit Ravi's Stanford RNA 3D Folding entry

Stanford RNA 3D Folding

Kaggle competition entry for predicting RNA 3D structure from sequence — a pipeline + evaluation stack using TM-align for structural scoring.

Illustration of a human painter and an AI robot painting the same scene — Rohit Ravi's AI vs. Human Art classifier project

AI vs. Human Art Classification

A deep-learning classifier that distinguishes AI-generated from human-made artwork, with interpretability via Grad-CAM, t-SNE, and PCA.

Pipeline diagram: Cellular Automata → Attention → Causal Graph → Symbolic Rules — architecture of Rohit Ravi's Hybrid Symbolic Discovery Engine

HSDE — Hybrid Symbolic Discovery Engine

Neural-symbolic reasoning prototypes: a recursive CA + attention + symbolic hybrid, and an Enhanced Symbolic Discovery Engine targeting AIME / SAT / GSM8K.

OnServe Tennis — Website

Website for OnServe Tennis, built in React. An earlier full-stack project from before I moved primarily into ML research.

Code-OCR

An OCR tool for interpreting source code images (.jpg, .png) and outputting machine-readable text — an early exploration of vision-to-code conversion.

GG Tennis Academy — Website

Website for GG Tennis Academy. An early web-development project from before I moved primarily into ML research.

Carbon Emissions Tracker — Florida Hacks with IBM

Hackathon project for Florida Hacks with IBM — a carbon emissions tracking app to help users quantify and reduce their footprint.