[Pranshu Rao][Projects][Reading][Writing]

Projects

A few things I've built, broken, and rebuilt.

[robotics / ml]
2026

clankeRS

A Rust SDK for robotics teams on ROS 2 and PyTorch: train in PyTorch, deploy ONNX inference in Rust, and replay-test nodes against real MCAP robot logs before they hit hardware.

RustPyTorchONNXMCAP
[view →]
[ml research]
2026

beyondOPD

Research on cheap methods for surpassing the teacher in on-policy distillation. Tries confidence-scaled (CSOPD) and verifier-gated hindsight (VGHOPD) objectives that reweight and outcome-gate distillation without extra rollouts or a task reward.

RLLLM Post TrainingPyTorch
[view →]
[chess engine]
2026

Beluga

A ~2880 Elo UCI chess engine written from scratch in Rust. Bitboard board representation with magic-bitboard move generation, an iterative-deepening PVS search with a transposition table and Lazy SMP threading, and a tapered handcrafted evaluation.

RustBitboardsLazy SMPNNUE
[view →]
[open source]
2026

ros2_rust

Open-source contributor to the Rust bindings for ROS 2. Added Clock sleep and async sleep APIs (sleep_for / sleep_until and their async variants) aligned with rclcpp semantics, so timers can yield to the executor under simulated time without blocking.

RustROS 2async
[view →]
[ai / games]
2025

Secrets

A cinematic, AI-powered social deduction game for the terminal — a 7-player Avalon spin-off where you play a lone human against LLM agents that argue, accuse, and scheme in real time. Powered by local Ollama models.

LLM AgentsMulti Agent Systems
[view →]
[rag / ml]
2025

liteRAG

A Next.js and FastAPI demo that builds tiny LiteRAG (LEANN) and Chroma indexes side by side, compares on-disk storage, and queries both — highlighting LiteRAG's smaller footprint with comparable retrieval on the demo dataset.

Next.jsFastAPILEANNChroma
[view →]
[deep learning]
2025

MNIST Neural Network from Scratch

A neural network implemented from scratch with only NumPy and pandas — no TensorFlow or PyTorch. Fully-connected layers with ReLU and softmax trained via backprop on MNIST, reaching ~94% test accuracy while making the underlying linear algebra explicit.

PythonNumPyML Theory
[view →]
[robotics]
2024

Kalman Filter for Autonomous Vehicle

A Kalman filter for sensor fusion on an autonomous RC car built by Formula Electric @ Berkeley. Fuses noisy sensor streams into an optimal state estimate for a linear-Gaussian system, with prediction/update steps and visualization of the tracked state.

PythonNumPyPandas
[view →]