Second-year Computing student at Imperial College London. I build robot perception systems and physical prototypes.
My current focus is coverage-based edge-case hunter for robot reliability: how to find failures that a good local benchmark can miss.
sienarindustries.com · LinkedIn · Hugging Face · Email
- Vikaso Robotics — Robotics & ML Intern. Built the company’s first calibration pipeline for a UR5e. Integrated two segmentation and four 6D-pose models behind one interface. Built a BOP evaluation harness that matched the models’ published Average Recall results.
- GT-GAP. Research on test coverage and failure discovery for learned manipulation policies. Cable insertion is the case study. Paper in preparation.
- Next. SO-101 sim-to-real experiments, an expanded task board, and tests of how simulation results predict physical performance.
-
Project-Ladder — CodeBase
Intrinsic’s AI for Industry Challenge, 49th of 196. Our UR5e cable-insertion policy scored 141 locally and 89 on hidden scenes. That gap led to my current work on test coverage and failure discovery. -
nanoGPT-Valkyrie — Research code · GPT-2 reproduction · Paper
GPT-2 124M reproduction and research on normalisation methods and their positions within Transformer blocks. Pre-trained models on four H100 GPUs with PyTorch DDP, gradient accumulation, and mixed precision. This was followed by downstream evaluation and ablation studies. -
Sparky / Hazel — Sparky · Hazel
A 2D engine and application framework in C++ and OpenGL, based on The Cherno’s series. Sparky includes a batched sprite renderer, GLSL pipeline, and hand-written matrix library. Hazel includes templated event dispatch, a layer stack, and Dear ImGui docking.
-
Augmented electromagnetic railgun — Write-up · Rail-launcher simulations
Three prototypes with a hand-wound ZVS transformer. Average projectile speed of 132 km/h across 10+ firings. A separate physics study used a rail-launcher model, with 70 LS-DYNA runs across 10 voltage conditions. -
FPV racing drone — Build log
Designed the PET-CF frame in Blender and soldered the electronics. Mamba F722, DJI O4 Pro, and ELRS; 610 g, tuned in Betaflight and flown in Acro mode. -
F1 in Schools — Anduril Racing — Team site
Team and technical lead. Took designs through Fusion 360, Ansys Fluent CFD, resin printing, and CNC manufacture. Built a 20 m test track.
- Languages: Python, C/C++, Kotlin, Java, Haskell, TypeScript/JavaScript, RISC-V assembly
- Machine learning: PyTorch, DDP, Hugging Face Transformers, LeRobot, YOLO, SAM, Weights & Biases, NumPy, vLLM, llama.cpp
- Robotics and simulation: ROS 2, MuJoCo C API, Gazebo, Isaac Sim, FoundationPose, OpenCV, Universal Robots UR5e
- Systems and graphics: Linux/CUDA toolchain, Git, Docker, GCP, AWS, OpenGL
- Web: React, Node.js, Firebase
- CAD and hardware: Fusion 360, Blender, Ansys Fluent, LS-DYNA, 3D printing, CNC, electronics, Betaflight
I run 25–30 km a week. Murakami’s writing helped me see endurance and consistent effort as skills I can practise.
I read science fiction as thought experiments about technology and power. I also study how major technological changes make established assumptions fail.
Books and authors: Liu Cixin · Frank Herbert’s Dune · Haruki Murakami’s What I Talk About When I Talk About Running