hello, i'm

Vedant Rathi

currently pursuing bachelor's in computer science at UIUC

passionate about deep learning, especially interpretability, post-training, and inference for LLMs.

Research

selected publications

Breaking Bad Tokens

Detoxifying LLMs using sparse autoencoders and interpretability-driven interventions. We decompose black-box models like GPT and Gemma to identify and suppress toxic features while preserving coherence.

EMNLP 2025

Teaching Reasoning to LLMs

Investigated inducing LLM reasoning without RL or distillation; contributed to a released human-authored chain-of-thought dataset for contest math, used for SFT.

ICML AI4Math 2025 Workshop

Interpretable Spam Detection

Meta-learner framework for email spam detection that combines ensemble ML and deep learning techniques. Evaluated on 50,000+ samples with strong zero-shot generalization.

Frontiers in AI 2025

Projects

things i've built

Tiny KV Cache

Embedded key-value store in C++20 with an append-only write-ahead log and on-disk snapshots. Implements thread-safe in-memory state via sharded hash maps with WAL replay for crash recovery.

  • C++
  • Threading
  • File I/O

Mini Inference Engine

Load-balanced, multi-container inference serving system for running small language models with real-time token streaming over gRPC. Routes requests using load balancer with live metrics and manages Dockerized worker fleets.

  • Docker
  • gRPC
  • Inference

Predictive Concept Decoders

Reproduction of Transluce AI's interpretability assistant. Trains an encoder to compress residual-stream activations into sparse concept bottlenecks, with a decoder that answers questions about model behavior.

  • PyTorch
  • DDP
  • Interpretability

CycleGAN Style Transfer

From-scratch CycleGAN implementation for unpaired image-to-image translation. Converts photos to Monet-style paintings and back, trained for 5 epochs on P100s.

  • PyTorch
  • CycleGAN
  • Computer Vision

LyricSearch

AI-powered lyrics platform that generates original verses with GPT-3.5 and checks authenticity against 480K+ real songs using RAG-style semantic search with FAISS.

  • RAG
  • FAISS
  • Flask

Gmail MCP Server

Model Context Protocol server that lets Claude interact with Gmail via natural language. Supports drafting, sending, labeling, and managing threads through conversational commands.

  • FastMCP
  • Gmail API

Experience

where i've worked

Fall 2025

ML Research Intern

Axiom

Worked on building multi-agent reasoning systems and improving RL post-training for formal theorem proving.

Summer 2025

ML Engineer Intern

MathWorks

Engineered a MATLAB coding assistant with a slash-command interface + MCP tool servers, implementing structured tool-calls (e.g., codebase search, doc retrieval, run/verify) for reliable multi-step workflows.

Spring 2025

Student Researcher

Nvidia Math Reasoning Team

Contributed to research on teaching reasoning capabilities to language models through minimal supervision approaches.

Summer 2024

ML Engineer Intern

AviaryAI (YC S22)

Fine-tuned LLMs to steer voice-agent persona consistency and style; built OCR-driven pipelines to extract and structure insurance documents for automation.

2024 - Present

Student Researcher

Crowd Dynamics Lab @ UIUC

Research on LLM safety and interpretability, focusing on methods to identify and mitigate harmful model behaviors.