Undergrad at Penn State — CS, Physics, AI, Schreyer Honors '28.
Working on mechanistic interpretability: reverse-engineering what language models compute, and why.

Writeups live at ameya-bit.github.io.

Building Anthropic's Jacobian lens from scratch — the method from Verbalizable Representations Form a Global Workspace in Language Models (Gurnee et al., 2026) — in raw PyTorch + HuggingFace, validated against finite differences and the logit-lens identity before any reference code is diffed. Reference implementations exist; the point is owning the mechanism, because the code becomes the probe for what comes next.

_{jacobian-lens-from-scratch · in progress, built in the open}

Every Method Has a Blind Spot: Reading GPT-2's IOI Circuit Three Ways
_{Jul 2026}

Direct logit attribution finds the name-mover heads, activation patching finds what feeds them, path patching finds the wire between — and each method is blind to exactly what the next one reveals. A replication of Wang et al. framed as a comparison of what the three techniques can and can't see, with DLA, activation patching, and path patching hand-implemented before reaching for library helpers.

_Code

An Induction Head in Disguise: Chasing Grammar in a Character-Level Transformer
_{Jul 2026}

Trained a 6-block character-level transformer, isolated a copying head (B5H0) via OV-circuit eigenvalues and copying scores, and killed two of my own hypotheses — quotation and parenthesis grammar — with targeted QK attention-pattern controls. The result: the OV circuit encodes bracket completion while the QK circuit stays blind to bracket structure. A half-built mechanism the model never learned to route.

_Code

Monte Carlo Radiative Transfer — Neutron-Star Pulse Profiles
_{Penn State Multi-Campus REU · advised by Prof. Asif ud-Doula · Oct 2025 – present}

A two-stage pipeline — a Monte Carlo scattering atmosphere feeding a general-relativistic pulse-profile model with Schwarzschild light-bending and Doppler effects — built to quantify the systematic error introduced by the isotropic-emission assumption in neutron-star mass–radius measurements. Validated against Chandrasekhar's exact H-function (reduced χ² = 0.70) and published reference waveforms (0.11% peak error). First-author paper in preparation.

_{Code · Project writeup}

CanvAI — semantic retrieval over Canvas LMS content; HuggingFace embeddings indexed in FAISS. PSU Hackathon, 2025.

Web vulnerabilities study — seven common vulnerability classes, with defensive implementations across Rust, Python, PowerShell, and PL/pgSQL.

Application Specialist (part-time) — Penn State Office of the Physical Plant, with the Nittany AI Advance.

Python · PyTorch · HuggingFace Transformers · TransformerLens · NumPy · scikit-learn · Git

_{Secondary: Java · JavaScript / React · SQL · Flask}

_{Website · LinkedIn · X · Email}