Computer science student working on machine learning and low-level systems — and fairly often on the overlap between the two. I'm drawn to problems where the algorithm and its implementation both matter: an index is only as good as its memory behaviour, and a model is only as useful as the pipeline feeding it.
Machine learning. Diffusion models and their conditioning mechanisms, transformer architectures for detection tasks, and reinforcement learning for control. I care about the experimental side — proper ablations, honest baselines, and results that survive scrutiny.
Systems and data structures. Storage engines, indexes and runtimes written from scratch: LSM trees, B+ trees, approximate nearest-neighbour search, Linux container primitives, compilers. Building these is how I learn what the abstractions actually cost.
Simulation and control. GPU-accelerated numerical simulation, classical and learned control strategies for robots, and streaming data pipelines.
Languages — Python, C++, C, Go, Java, JavaScript ML — PyTorch, scikit-learn, Hugging Face Transformers, OpenCV, CUDA Data & infrastructure — Spark, Kafka, Elasticsearch, MySQL, Docker, Kubernetes
B-PLUS-TREE — Hybrid B+ tree and HNSW index for k-NN search with scalar filtering. A probabilistic selection strategy keeps recall high when filters would otherwise starve the graph traversal. C++
Persistent-KV-Store-Engine-in-Golang — Persistent key-value store built on an LSM tree: write-ahead logging for durability, SST compaction to bound read amplification, and concurrent access through goroutines. Go
Cool-Compiler — Complete compiler for the Cool language, from lexing and parsing through semantic analysis to LLVM IR generation with optimization passes. Go
cloud-minic — Mini container runtime built directly on Linux primitives: PID/UTS/mount/IPC namespaces, OverlayFS with pivot_root, and cgroups v2 resource limits, with a Flask dashboard for monitoring. Python
Controlled-Conditioning — Experimental comparison of ControlNet, LoRA adapters and input concatenation as conditioning mechanisms in diffusion models, with rank and layer ablations testing whether conditioning signals are genuinely low-rank. PyTorch
RayleighTaylor-Cuda — CUDA-accelerated CFD simulation of the Rayleigh–Taylor instability using the 2D compressible Euler equations. Companion to MaxReduction-Cuda, a six-stage study of parallel reduction from naive global memory to warp shuffles. CUDA
mini-sklearn — core ML algorithms, preprocessing and metrics reimplemented from scratch · Robotics-project — classical and RL controllers compared across three robot models · Deepfakes — transformer-based detection for image and audio deepfakes · Real-Time-Financial-Data-Analytics — Kafka to Spark to Elasticsearch streaming pipeline · GRAPH-VI — D3.js graph explorer for malware sandbox reports
Conditioning mechanisms in diffusion models, and comparing classical against learned control strategies for mobile robots and manipulators.