Technical founder.
Builder and researcher.

I built a reasoning-aware text-to-image system that gives diffusion models an explicit intermediate representation before generation. Instead of treating a prompt as one fuzzy embedding, I designed an enriched scene graph that captures objects, counts, attributes, relationships, and reasoning constraints. I also built an image-editing chatbot where the scene graph acts as editable state: when a user asks to “add a window” or “make the couch green,” the graph updates directly and the image regenerates. This made image editing more programmable and controllable than repeatedly prompting a black-box diffusion model.

• Diffusion Models
• Scene Graphs
• LLMs
• Image Editing
• Structured Reasoning
• Python
• PyTorch

I built an optimizing L4 compiler in Rust end to end: parsing and elaboration, typechecking with a string interner, lowering to an SSA IR tree, translation to LLVM-like SSA, 3-address x86 SSA, and a backend with dominator analysis, if-conversion, graph-coloring register allocation, SSA deconstruction, and AT&T x86 emission. On the optimization side, I implemented early IR cleanup (copy and constant propagation, folding, DCE), repeated LLVM scalar passes (CSE, arithmetic strength reduction, branch folding), function inlining, tail-call elimination, LICM, partial redundancy elimination, induction-variable strength reduction, scalar replacement, and final x86 peepholes. I tuned the pipeline with per-pass ablations on hot benchmarks. Our O1 pipeline reaches over 70% of optimized GCC.

• Rust
• L4
• Compilers
• SSA
• Graph Algorithms
• x86
• Optimization

We built a minimal PyTorch-like stack entirely on Needle, including tensors and autograd, CPU and CUDA NDArray backends, neural modules (Linear, Conv, activations), optimizers, and CIFAR-10 training with a ResNet-style model. Every layer runs on our own framework rather than PyTorch or TensorFlow. We then implemented TVM-inspired lazy graph compiler passes: constant folding, elementwise chain fusion with a single CUDA interpreter kernel, and matmul, bias, and activation epilogue fusion (tiled GEMM plus activation in registers), plus BroadcastLike so nn.Linear stays lazy. Benchmarks show fewer graph nodes, fewer kernel launches, and faster training versus an unfused baseline.

• Python
• CUDA
• Needle
• Autograd
• Graph Optimization
• Deep Learning Systems

I built a hybrid quantum-classical classifier for malignant breast cancer diagnosis. Instead of only training a normal neural network, I encoded patient features into a quantum circuit and tested whether that representation could learn a better decision boundary. I compared classical-only, quantum-only, and hybrid PennyLane/PyTorch models, and the hybrid model improved test accuracy by about 2 percentage points over the classical baseline.

• PennyLane
• PyTorch
• Quantum ML
• Biomedical Classification
• Hybrid Models
• Model Evaluation

We built Lighthouse, a Chrome extension that captures periodic screen frames through Chrome's offscreen display-media pipeline and streams them to a Flask backend. On each frame, we combine geometry-aware OCR (center- and size-weighted Tesseract bounding boxes via OpenCV) with GPT-4o structured vision outputs such as app or site, activity summary, and salient on-screen objects. A schema-constrained classifier scores distraction against user-defined focus and anti-goals from zero to four, then the extension escalates lock-in notifications and audio nudges in real time. We also built an offline evaluation loop with scraped screenshots and Google's Screen QA dataset to stress-test the pipeline before demoing live.

• Python
• Chrome Extension
• Flask
• Computer Vision
• GPT-4o
• OCR
• Real-Time Systems