Chinmay Khanna / AI Engineer / LA

I'm a CS master's student at USC researching self-learning agent harnesses in game environments, building a harness that lets an AI agent plan and act across 10 GameBoy titles, including Pokémon, Zelda, Harry Potter, Deja Vu and Ruins of Virtue without any prior knowledge. I'm also building AI products at Avis Budget Group, working against the scale of an operation that spans 180+ countries.

Outside that I just like learning things, exploring new places and cooking. This semester at USC I'm taking Robot Learning and Autonomous Cyber-Physical Systems.

500-task embodied agent benchmark, under review at ICLR 2027
3.9/4.0 undergrad GPA, top 5% of cohort
AI Engineering at Avis Budget Group — 180+ country operation
Published & presented at IEEE ISCS 2024
Research Assistant at USC Information Sciences Institute
Beat Mistral-7B on grammatical error correction at ~25× fewer parameters

Experience

Built AI products for the parent of Avis, Budget, Zipcar, Payless and Apex, operating across 180+ countries. Most work began as a problem statement rather than a spec, so it started with sitting with each team to learn what they actually measured, then building the full-stack tool (internal sites, MCP servers) across marketing, operations and the field, with senior leadership in the room.

Parsippany, NJ

Researching self-learning agents and autonomous skill discovery. A testbed of 500 tasks across ten Game Boy games where the agent gets no reward and no demonstrations, plus a three-tier strategist–supervisor–executor harness that breaks goals into subgoals and plays through titles it has never seen, building its own knowledge base as it goes.

Gemini successfully exiting a cave

Los Angeles, CA

Grader for EE-467, Introduction to Communication Systems.

Los Angeles, CA

Co-founded a freelance studio delivering software for external clients over two years.

Remote

Worked with the R&D team on pipelines for AI-generated content — 20+ shows, five openings each. Every opening shipped as a real ad and was growth-tested on Meta, so the audience decided which version of a show got developed before anyone committed to making it.

This is how dedicated an employee I am

New Delhi, India

lately

hover to pause, drag to spin, click to open

Featured Projects

GameBoyWorlds

Research · ICLR 2027 under review2026

A testbed for self-improvement in embodied game environments. Ten Game Boy titles across five series, modelled as POMDPs with a discrete eight-button action space, 50 hand-curated tasks and completion functions each, for 500 tasks total. Agents train with no rewards, no demonstrations and no documentation, then are dropped into a moment of gameplay and given a task in natural language.

500 tasks · 10 titles · 5 series
Best frontier model averages 45%
Deja Vu games near-unsolved at 2–22%
Open Gemma-4-31B (35%) beats GPT-5-mini (26%)

AV1 Simulator

CSCI 576 · Multimedia Systems Design2025—26

An end-to-end AV1-inspired video codec: raw YUV420 in, AV1S bitstream out, fully decoded back to playable video. Adaptive block partitioning sizes blocks by local detail, then spatial and temporal prediction, transform coding, quantization and entropy coding each stage visible and inspectable. Wrapped in a streaming-service demo UI that runs the pipeline on click.

12.7× compression with preserved reconstruction
Variable block sizes driven by edge detail
Full encode → bitstream → decode round trip
Every pipeline stage independently visualized

Repo/Live demo

Multilingual Grammatical Error Correction

CSCI 544 · Applied NLP2025

Two XLM-RoBERTa-based approaches to the same problem: one learns which edits to apply to each token, the other writes the corrected sentence from scratch. Reframing correction as token-level keep/delete/replace/insert edits applied by a deterministic rewriter — rather than generation — is what beat a frontier model many times its size.

0.91 F₀.₅ English / 0.88 Spanish
Beats zero-shot Mistral-7B (0.41 / 0.39)
~25× fewer parameters than the baseline
105K English + Spanish sentence pairs

Repo

Handin

Local-first desktop app2026

A desktop workspace for notes, task lists and a calendar that stores everything as plain Markdown under ~/handin and backs it up to a private repo on your own GitHub account. No account with us, no cloud database holding your writing hostage. Open the folder in Finder and it is all there — readable, greppable, yours.

Atomic writes — a crash mid-save never truncates a note
isomorphic-git built in, no system Git dependency
Divergent history saved to conflicts/ before adopting remote
Telegram → Supabase inbox with 1-hour signed URLs

Site

By the Numbers

Benchmark tasks authored across 10 Game Boy titles: 500Benchmark tasks authored across 10 Game Boy titles
Best frontier model's mean task completion on GameBoyWorlds: 45%Best frontier model's mean task completion on GameBoyWorlds
SHA-256 content blocks indexed from 10.8K avis.com pages: 545KSHA-256 content blocks indexed from 10.8K avis.com pages
Kickstarter campaigns scraped and modelled to predict which factors contribute most to a campaign's success: 23,990Kickstarter campaigns scraped and modelled to predict which factors contribute most to a campaign's success
Compression ratio from the AV1 pipeline: 12.7×Compression ratio from the AV1 pipeline
Increase in daily promo video throughput at Pocket FM: 5.3×Increase in daily promo video throughput at Pocket FM
Countries in the operation I build tooling for at Avis Budget Group: 180+Countries in the operation I build tooling for at Avis Budget Group
Container image reduction via multi-stage builds at avis.com: 1.3GB→310MBContainer image reduction via multi-stage builds at avis.com

Other Projects

Predicted campaign success and fraud risk across 23,990 Kickstarter campaigns scraped with Selenium, engineering deception-cue, grammatical-error and Benford's Law features to reach 0.899 accuracy with logistic regression. Presented at the IEEE International Conference on Intelligent Systems for Cybersecurity.

Paper

Reconstructs images from randomly rotated and translated fragments, regular and irregular. Matches complementary edges by Mahalanobis Gradient Compatibility over Sobel gradient distributions, then recovers placements with beam search over a 5,000-state frontier, outputting the mosaic as a step-by-step animation.

Repo

Hides text, images or audio inside WAV files via least-significant-bit encoding, shipped as a Dockerized Flask API with S3 storage and local fallback, a web client, and a SwiftUI iOS app. Includes waveform visualizations of how the carrier signal changes after embedding.

Repo/Live

A grounded retrieval-augmented chatbot over a structured knowledge base, using hybrid search to answer with citations, confidence scores and explicit insufficient-context handling rather than confabulating — served through FastAPI with a local web UI.

Repo

Neural networks learning to protect a channel against a learned adversary — two models converge on an encoding scheme while a third is trained to break it.

Repo

Education

University of Southern California

Aug 2025 — May 2027

M.S. in Computer Science

Grader, EE-467
Research Assistant, USC ISI

CSCI 642Robot LearningCSCI 513Autonomous Cyber-Physical SystemsCSCI 576Multimedia Systems DesignCSCI 544Applied Natural Language ProcessingCSCI 570Analysis of Algorithms

Guru Gobind Singh Indraprastha University

Aug 2021 — Jun 2025

B.Tech. in Information Technology, AI/ML Minor

3.9 / 4.0 GPA — top 5% of cohort
AI/ML Lead, Google Developer Student Club
Tech Head, TEDxBVCOE
Winner, VesHack
Best Use of AssemblyAI API, BC Hacks 3.0
Presenter, IEEE ISCS 2024

Data StructuresObject-Oriented ProgrammingDigital Signal ProcessingMachine LearningStatistics

Skills

001

Languages

Python, C++17, Java, JavaScript, TypeScript, Swift, SQL. Comfortable dropping to C++ when the pipeline stops being fast enough in Python.

002

ML & Systems

PyTorch, TensorFlow, scikit-learn, OpenCV, vLLM, LangGraph, Hugging Face Transformers, CUDA. Self-hosted a 31B model on a SLURM cluster for benchmarking.

003