Naval Agarwal

CSE (AI/ML) · VIT Vellore  ·  Autonomous software at Team Ojas Racing

About  ·  What I'm working on  ·  Selected projects  ·  Stack  ·  Elsewhere

I'm a second-year CSE (AI/ML) student at VIT Vellore. Most of my time goes into the autonomous software for our Formula Student driverless car at Team Ojas Racing — the perception, planning and control that has to run in real time, on a real car, with nobody touching the wheel.

Away from the team I build a fairly wide mix of things: an LLM robustness research framework, a few C++ decision engines, some security tooling, and the occasional web app. The common thread is that I like problems where the program has to actually decide something — pick a racing line, flag a risky user, choose which loan to pay off first — and where being wrong has a real cost.

Team Ojas Racing — Formula Student Driverless (EV). Autonomous Software Developer, and Operations / CR on the side.

I work across the autonomy stack on a ROS2 Jazzy architecture: stereo depth from an Intel RealSense D435i fused with an SBG Ellipse N IMU for perception, the planning and control that turns that into steering and throttle, and getting all of it to run on a Jetson Orin Nano on the car. The interesting part isn't making it work once — it's making it correct enough to trust at speed.

Tighter sensor fusion and SLAM, getting perception to run faster on the Jetson (CUDA / TensorRT), and reinforcement learning for control rather than just for games.

Autonomy & robotics

• TOR27 — FSD Path Planning · Python A path-planning pipeline for the driverless car, with a live viewer so you can watch it think. It maps the track from cone detections, plans a line through it, and drives all four Formula Student missions — acceleration, skidpad, autocross and trackdrive. It also models the safety side: the AS state machine, the ASMS / RES / R2D arming sequence, and emergency braking. Still being validated on hardware, so it's honestly work in progress.

• speed-racer-rl · C++ A custom 2D racing simulator with a Deep Q-Network agent (libtorch). The car perceives the track through 13 LIDAR-style raycasts (an 18-dimensional state), trains headless, and trained models can be replayed visually with Raylib.

• SBG Ellipse N IMU nodes · Python · ROS Three ROS nodes — publisher, processing and telemetry — for the SBG Ellipse N IMU that feeds the car's autonomous stack.

ML & AI research

• G0DM0D3 · TypeScript A modular framework for studying how large language models behave at inference time, without any access to their weights. Four composable pieces: AutoTune (context-aware sampling), Parseltongue (character-level input perturbation for red-teaming), STM (output normalization), and ULTRAPLINIAN, which races up to 51 models in parallel and scores them on one metric — plus an EMA feedback loop that learns from ratings. ~3,300 lines of TypeScript, with the full write-up in PAPER.md.

• ML-Assisted Adaptive Cache · ML · Systems Most CPUs run one fixed cache configuration no matter what the program is doing. This is an attempt to learn the workload instead — watch hit/miss rate, IPC and memory access frequency at runtime, then predict a better cache configuration for the current phase.

Web & tools

• CodeLens · TypeScript · Next.js · source Paste a GitHub repo URL and get a code-quality score with a breakdown — complexity, maintainability, style and bug safety — across 15+ languages. Live on Vercel.

Security

• Insider Fraud Detection · Python Endpoint and network monitoring for insider risk: device discovery, per-device traffic, USB events and remote agents, feeding an Isolation Forest model that scores users — and uses SHAP to explain why someone got flagged, including Big-Five behavioural signals. All of it behind a dashboard.

Systems & embedded

• Adaptive Loan Repayment Scheduler · C++ Give it a set of loans and a monthly budget; it recomputes repayment priority every cycle from interest, tenure and balance, and returns the cheapest order to clear them.

• Tremor Stabilization Device · C++ · Arduino An assistive device that keeps a spoon or brush steady for people with Parkinson's or essential tremor. An MPU6050 reads the tremor, an Arduino Nano classifies it, and a servo applies counterforce in real time.

Reach me on LinkedIn or at navalag123@gmail.com.