RBmove your cursor through it · click to morph

Ruchir Bhatia / Software & AI Engineer / LA

I build software that has to be right at 300 km/h.

I'm a CS master's student at USC who ships things end-to-end: a live F1 race-strategy engine that makes pit calls from real timing data, an on-device driving co-pilot that watches the road and the driver, and customer-facing features for a product that doubled its users while I was on the team.

I care about one thing more than most: numbers I can defend. Every model on this page was tested on data it never saw, with the test written down before it was run. Where something didn't work, I say so.

📍 Los Angeles, CA🎓 USC M.S. CS · May 2027

Experience 01

Where I've shipped code other people depend on.

Transnode AI Software Engineer Intern · letter ↗Jun – Aug 2026

Owned customer-facing features for Earnie end-to-end across React Native, iOS and web, from design through launch, including the release-blocking bugs nobody else wanted.

679 → 1,436 registered users (~111%)React Native · iOS · Web
Bharati Vidyapeeth COE ML Research AssistantJan – Jun 2024

Built a retrieval-augmented generation pipeline grounding LLM answers in structured knowledge, then profiled it until it was fast enough to feel instant.

65% lower end-to-end latencyPython · PyTorch · RAG
Redis Ltd Software Engineer Intern · letter ↗Jun – Sep 2023

Designed and tested Redis 7.2 caching workflows for high-throughput, low-latency access on AWS ElastiCache, and automated the regression suite around them.

18% less memory30% faster testing

Featured Projects 02

Real recordings of the real systems, and one you can drive yourself.

PitWall's live pit wall replaying the 2026 Azerbaijan GP lap by lap through the same engine that runs on race day: "stay out" on lap 13, "BOX, BOX, Softs" on lap 29, then "stay out to the flag". Independent project, not affiliated with Formula 1.

PitWall AI

2026 · deployed

A live Formula 1 race-strategy engine. A Bayesian tyre-wear model updates lap by lap from live timing; an exact dynamic-programming optimiser searches every 0–3 stop plan; Monte Carlo simulates safety-car risk. It makes a pit call in under three seconds.

Walk-forward on 173 driver-racesFirst-stop error 9.2 vs 10.3 laps (baseline)Stop-count accuracy 33.5% vs 27.7%

Honest note: the baseline is a naive strategy, and the margin is real but modest. F1 strategy is hard, and that's why I built it.

FastAPINext.jsBayesian inferenceDynamic programmingMonte Carlo
Live app ↗Repo ↗

Screen recording of DriveMind running live on a MacBook Pro, everything on-device. Road footage: BDD100K © 2018 The Regents of the University of California, used for non-commercial, educational purposes under the BDD100K license.

DriveMind: Multimodal In-Vehicle Assistant

2026

An on-device co-pilot: YOLO11 + ByteTrack perception on the road, a driver monitor watching for drowsiness, and a 4-bit vision-language model agent you can talk to through Whisper.

Eye-state model blink F1 0.95False microsleep alerts 33/h → 010/12 unseen drowsy drivers caughtYOLO 10–18 ms/frame
Spoken question 'Is that light green?' answered on-device: Whisper 0.6 s, VLM 1.7 s, 2.3 s end to end
Ask it out loud, “Is that light green?”, and it answers on-device in about 2.3 s (Whisper 0.6 s + VLM 1.7 s).

Every drowsiness test was pre-registered and run once on drivers the system had never seen. Two earlier traffic-light tests failed, and they're in the repo too.

PyTorchYOLO11MediaPipeWhisperVLM agent
Repo ↗How it was tested ↗

Sociopals

fullstack · solo

A real-time social platform designed and shipped end-to-end: authentication, access control, live messaging, and audio/video channels, on reusable APIs and a relational data model.

Next.jsTypeScriptNode.jsPrisma / MySQLSocket.io
Repo ↗

By the Numbers 03

Measured, not estimated.

0user growth at Transnode AI while I shipped features
0lower latency on a RAG pipeline
0driver-races in PitWall's walk-forward backtest
<3sfrom live timing to a pit call
10/12unseen drowsy drivers caught by DriveMind
0.95blink F1, leave-one-person-out
0memory saved on Redis ElastiCache
0publications: IEEE paper, patent, research paper

Publications 04

 

IEEE · Published
OS Scheduling & Page Replacement

Peer-reviewed analysis of page-replacement and CPU-scheduling algorithms for OS memory management.

Patent · Accepted
LLM Healthcare Chatbot

Retrieval-augmented chatbot grounded in structured medical knowledge for clinical question answering.

Research Paper
VLM4Rec: Multimodal Recommendation

Grounds item images into semantic descriptions with a vision-language model, outperforming fusion-based baselines.

Software Copyright · Registered
Page-Replacement Algorithms in Shell

Education 05

 

University of Southern California M.S. Computer Science– May 2027
CSCI 566 Deep LearningCSCI 570 AlgorithmsCSCI 642 Robot LearningDSCI 552 ML: 4.0, joint top scorer
Bharati Vidyapeeth College of Engineering B.Tech Computer Engineering– Jun 2024
GPA 3.87 / 4.00Top 5% of classAlgorithmsOperating SystemsMachine Learning

Skills 06

 

001Languages

Python, C++, TypeScript / JavaScript, SQL, Swift, Bash

002ML & AI

PyTorch, TensorFlow, LLMs, RAG, vision-language models, Whisper, YOLO, MediaPipe, XGBoost / LightGBM, honest evaluation

003Software

React, React Native, Next.js, Node.js, FastAPI, REST APIs, real-time systems, system design

004Cloud & Data

AWS (EC2, S3, ElastiCache), Redis, MySQL, Docker, Vercel, Render, Git

Let's build something.

I'm looking for full-time Software, ML and AI engineering roles starting mid-2027. The fastest way to reach me is email. I reply within a day.

✉ bhatiaruchir04@gmail.com LinkedIn ↗