Rayhan Khilji

TensorVM is a small neural network runtime built from scratch in Rust to understand and implement the systems underneath modern ML frameworks. The runtime represents models as a tensor intermediate representation (IR), with the same computation graph used for forward execution, reverse-mode automatic differentiation, model serialization, and backend execution. I implemented the core tensor system, computation graph, autodiff engine, backend abstraction, portable model format, and Python API. The core runtime is written in Rust and exposed to Python through PyO3, with a PyTorch-style frontend for defining models and optimizers. TensorVM currently supports CPU execution through a readable reference backend and an optimized backend, where blocked matrix multiplication delivers approximately 4.4x to 6.3x higher throughput on benchmarked matrix sizes. I also built end-to-end MNIST training and serialization, allowing a trained model to be saved as a portable `.tvm` file, loaded without the original model code, and executed on a different backend while producing identical predictions. The project is deliberately small and legible rather than production-scale. The goal was to understand how tensor runtimes, computational graphs, automatic differentiation, execution backends, and model serialization work underneath abstractions such as PyTorch and TVM. Tech: Rust, Python, PyO3, NumPy, ndarray, automatic differentiation, tensor IR, compiler/runtime design, numerical computing, machine learning systems. Open source : https://www.github.com/rayhankhilji/tensorvm

An online version of monopoly where you can play against bots of different skill levels. Built custom frameworks to solve edge cases like trading properties, mortgage and remortgaging, jail and free parking. Engineered to be fast paced, engaging and responsive, so instead of 3 hour long games, each game takes around 25 minutes and is just as fun. I am still working on the multiplayer mode. The project is open source. You can view the code here and contribute: https://github.com/rayhankhilji/parkway

Personalised horror game built with realtime ai world model technology, generating the game frame by frame based on refined inputs from user interview powered by openai, reactors technology and veed's interactive models.

Built a voice-first language learning platform designed around real conversations rather than flashcards or grammar exercises. Users practice speaking in realistic scenarios including job interviews, client meetings, ordering food, networking events and everyday conversations. The AI adapts the difficulty, vocabulary and speaking speed in real time based on the user's fluency, confidence and mistakes, while remembering previous sessions to build long-term progress. Implemented a low-latency conversational pipeline using streaming speech recognition, LLM reasoning and neural text-to-speech to create natural back-and-forth dialogue with minimal delay. Conversation history, vocabulary progress and grammar weaknesses are persisted in PostgreSQL, allowing lessons to build on previous sessions instead of starting from scratch. Built using TypeScript, React, Node.js, PostgreSQL and Redis with Groq Whisper for transcription, Gemini for conversational reasoning and ElevenLabs for speech synthesis. Additional features include pronunciation analysis, vocabulary revision, contextual grammar feedback, role-specific learning paths and simulated real-world conversations. Focused on helping users become conversationally fluent by simulating situations they are likely to encounter in real life, replacing repetitive language exercises with interactive spoken practice. Download for free here: https://voluble.works

AS 2026: ADU. GCSEs (2025): AAAAAAAABCC across Maths, English Lit/Lang, Geography, French, Biology, Chemistry, Physics, Design & Technology, Philosophy, Psychology.

2nd place in National Finals! 3x member Built Voluble (https://www.voluble.works) Built Lumi (https://www.trylumi.co) Building sidequest (https://www.sidequest.cyou)

Activities and societies: Language Captain Y10, Y11