Chinmay Rawat — The Lab

Dot plot of retained phase signal R for Gemma-2 2B and 9B across three ablation rungs. Marked traces sit high (0.63–0.78); with markers removed both models collapse to about 0.28–0.34, and the base–IT gap intervals cross zero. — The headline result. Retained signal R at fixed relative-depth layers across the decorrelation ladder, with 95% paired-bootstrap intervals and the frozen illusion / partial / abstract bands. Removing markers roughly halves retention at both scales.
Paper · resources/116_When_ReAct_Phase_Probes_Lo.pdf p.5, Figure 2

01

When ReAct Phase Probes Look Abstract

Mechanistic Interpretability ResearcherWork in progressApr 2026 – Present

A preregistered test of whether a linear probe that decodes ReAct reasoning phase at 98% accuracy is reading an abstract reasoning state or a correlated surface cue.

mechanistic-interpretabilityprobingsparse-autoencodersgemma-2preregistrationtransformerlens

AI Interpretability @ Illinois — group ↗

Calibration stress-test table on 306 tasks. The cross-fit calibrated prior has the best Brier score (0.216); the grounded LLM hybrid ranks similarly but has worse Brier (0.235). — The calibration stress test. A plain cross-fitted calibrated prior (Brier 0.216) beats the grounded LLM hybrid (0.235) — the result that fails the predictive-validity gate.
Paper · resources/36_More_Context_Is_Not_Validat (1).pdf p.3, Table 2

02

More Context Is Not Validation

AI Researcher Intern, Production LLM Systems & Evaluation InfrastructureWork in progressMay 2026 – Present

A three-gate audit of whether LLM agents that simulate individual listeners actually predict them, or only sound like they do.

llm-evaluationcalibrationagent-simulationbaselinesconstruct-validitybootstrap

ME

03

Medical Policy Digital Twin

Tech Projects Lead, Technology Partner (4 concurrent tech companies) — engineering lead on this projectOTCR ConsultingAug 2025 – Present

A policy simulator that replays a payer’s historical claims against proposed policy edits and forecasts the cost impact before the policy ships.

healthcaresimulationetlpostgresqlawsstreamlit

04

Predicting Internal Migration in the Brazilian Amazon

ML Lead & SPIN Research Fellow, Distributed Training & Data InfrastructureNCSA Research Conference 2026 (poster)Aug 2025 – Present

A graph neural network that predicts internal migration in the Brazilian Amazon as directed flows between municipality pairs — paired with gradient-boosted trees and SHAP for feature-level explanation across 5,570 municipalities and 300K+ town-to-town flows.

graph-neural-networksgraphsage-attentionxgboostshapgeospatialmigration

NCSA SPIN — 2025-2026 Academic Year Interns ↗NCSA SPIN program ↗

Line chart of linear-probe accuracy versus layer for GPT-2 Small, near 0.93 at every layer against a 0.25 majority-class baseline — an almost flat curve. — Tier 2. The probe hits ~93% at every layer against a 25% baseline. The README reads that flatness as a warning sign, not a finding.
Repository · github.com/Chinmayrawat15/ReAct_interp — figures/tier2_probe_accuracy.png

05

ReAct Interp — the GPT-2 pilot

Sole committer on the public repositoryApr 2026 (repository created and last pushed 2026-04-28)

A one-week interpretability sprint on GPT-2 Small that found ReAct phase trivially decodable at every layer — and said so, listing the controls it had not run.

mechanistic-interpretabilitygpt-2sparse-autoencodersprobingexploratorypython

Repository ↗README ↗

Three choropleth maps of Mexico by municipality from the paper: mean annual temperature (reds), total precipitation (blues), and GDP per capita (greens to blues). — Figure 1 from the paper — mean temperature, precipitation, and GDP per capita across Mexican municipalities. Click to enlarge.
Source · PAPER-C p.2, Figure 1 — Beyond Gravity Models

06

Beyond Gravity Models

Co-authorUniversity of Illinois Urbana-Champaign (NSF REU)2026

A co-authored paper on a graph neural network for internal migration in Mexico that beats gravity-model baselines by adding a spatial gate and an income-conditioned climate gate — with a built-in interpretability boost.

graph-neural-networksmigrationgravity-modelsspatial-lagclimateinterpretability

Two of the upstream projects contributed to. Logos are the trademarks of their respective projects and appear here to identify them, not to imply endorsement.

Source · Project logos (provided) — OpenClaw and PyTorch

07

Open Source Contributions

Open Source Contributor, ML frameworks & agent toolingJan 2026 – Sep 2026 (dates of first and most recent pull requests found)

Merged bug fixes to TransformerLens, marimo, OpenClaw and pylint, plus open pull requests to PyTorch, LiteLLM, Prefect, vLLM and Hermes.

open-sourcetransformerlenspytorchmarimoopenclawlitellm

All pull requests by Chinmayrawat15 ↗TransformerLens #1649 (merged) ↗

The live Phantom demo. Intended for authorised testing of targets you own or have permission to test.

Web · https://phantom-hackillinois.pages.dev (live site, retrieved 2026-08-26)

08

Phantom

Team Lead (team HackVoyagers)HackIllinois 2026 · YC Summer School 20262026 (repository commits Feb–Mar 2026)

An autonomous penetration-testing agent: point it at a target and it crawls the attack surface, plans and executes multi-step exploit chains in a ReAct loop, and returns CVSS-scored findings with reproduction steps in under 60 seconds.

securityagentsreact-loopplaywrightcloudflare-workersmodal

Live demo ↗

Marketing hero for "Copilot by PocketFM": a hand holding a stylus over a dark background, with the headline "Write smart. Write fast. Go global." and the subtitle "Adapt your story to 10+ languages, collaborate with other writers, and do more with our AI-powered Copilot." — PocketFM’s public "Copilot" writing product, which advertises adapting stories to 10+ languages.
Source · PocketFM — Copilot product marketing image ("COPILOT by PocketFM")

09

Adaptation

AI Researcher Intern, Production LLM Systems & Evaluation InfrastructurePocketFMMay 2026 – Present

A production LLM service that rewrites full scripts into a target language, adapting cultural references rather than translating them.

llm-applicationlocalizationproductionnlp

Three phone mockups of the PocketFM audio-series app on a magenta background, showing an episode player, a "My Vampire System" fantasy title with 67M+ plays, and a "God Eye" series page with play counts and ratings. — PocketFM’s audio-series app — the catalogue whose launch outcomes the Blockbuster Engine is built to forecast. Illustrative product image, not the engine itself.
Source · PocketFM — app / product marketing image

10

Blockbuster Engine

Software Engineer Intern, GenAIPocketFM (Bangalore, India)May 2025 – Jul 2025

An ensemble of fine-tuned LLMs and tree/neural models predicting content launch outcomes, deployed as containerised microservices.

loraensemblesgradient-boostingmicroservicesdockerproduction-ml

BeClear — brand wordmark. Logo supplied by the founder; no public product page was found.

Source · BeClear — brand logo (provided)

11

BeClear

Co-Founder2024 – 2025

An AI college-admissions agent, built on a RAG pipeline with LoRA fine-tuning; sold in 2025 to a group of schools in India.

raglorastartupedtechllm-application

MU

12

Multi-Agent Systems & Security

Researcher, Multi-Agent Systems & SecurityAlta Research Group, UIUC (advisor: Prof. Hao Peng)Aug 2026 – Present

Two connected questions in multi-agent coding safety, with Prof. Hao Peng: whether sub-agents can covertly pass confidential data past an LLM monitor, and whether the reasoning agents write down is what actually drives what they do.

multi-agentai-safetysecurityevaluationllm-monitoringreasoning-faithfulness

Hao Peng — Siebel School faculty page ↗

TE

13

Test-Time Scaling for Coding Agents

Researcher — Agentic Systems (with Prof. Hao Peng)Alta Research Group, UIUC (advisor: Prof. Hao Peng)Sep 2026 – Present

An early system that lets coding agents explore a codebase over many interactions to surface new, reusable knowledge about a task — with exploration pushed into the background so it stays cheap for users.

agentscoding-agentstest-time-scalingmulti-agentorchestrationearly-stage