Projects
Systems work across ML inference, GPU kernels, developer tools, market data, and infrastructure.
Showing 14 of 14 projects.
Featured Systems Work
Selected work in this project group.
Improved MoE forward-pass speed by approximately 1.3 times against the ragged-dot baseline.
Built and trained an approximately 10M-parameter decoder-only Transformer from scratch in JAX, Flax, and Optax for algorithmic reasoning. Extended it with a top-1 MoE layer and a custom fused Pallas projection kernel, achieving approximately 1.3x faster MoE forward passes than the ragged-dot baseline.
Raised throughput by 1.3 to 1.6 times and lowered latency by 15 to 20 percent.
Reproduced autoregressive, standard speculative, and Speculative2 decoding with Qwen3-32B and Qwen3-8B. Built a custom draft-verify runtime and benchmarked batches 1-64, reaching 1.3-1.6x higher throughput and 15-20% lower latency while identifying diminishing returns beyond speculative depth 4.
Reduced matmul latency by 10 to 13 percent and changed decode complexity from O(T^2) to O(T).
Profiled GPT-2 inference on an NVIDIA A40 and found matmul consumed 97-99% of GPU time at only 1-4% of peak compute. Reduced matmul latency by 10-13% using higher CTA counts, shared-memory tiling, loop unrolling, and vectorized loads; implemented WMMA GEMM and KV caching to reduce autoregressive decode complexity from O(T^2) to O(T).
Supported more than 1,200 AI-assisted edits for more than 50 alpha users.
Built an agentic IDE on a forked VS Code codebase with multi-agent execution, patch review, incremental indexing, vector search, live architecture graphs, Merkle-tree change detection, and an observability timeline. Onboarded 50+ alpha users and supported 1,200+ AI-assisted edits, with more than 60% reporting faster understanding and debugging.
More Projects
Earlier product, systems, and infrastructure work.
Unified visual agent composition, persistent memory, modular tools, and live streaming in one platform.
Created a full-stack platform for composing AI agents and workflows from text prompts with a visual editor, persistent memory, and session management. Built a modular tool ecosystem and a WebSocket streaming engine for live responses.
Connected secure ingestion, automated transcoding, and global delivery across a microservices architecture.
Built a microservices video platform with secure ingestion, automated FFmpeg transcoding, and global CloudFront delivery. Deployed the services through Docker and Kubernetes on AWS EKS with RabbitMQ message queues.
Connected packet decoding, columnar storage, orchestration, and query bookkeeping in one data path.
Built a high-performance C++ parser for IEX PCAP market data. Added Parquet storage, a SQLite bookkeeper, multithreaded ingestion, and Apache Airflow orchestration.
Enabled conflict-free multi-user editing with live cursors and document versioning.
Developed a real-time collaborative editor with Yjs CRDT synchronization, custom WebSocket infrastructure, and an AWS serverless deployment. Added live cursors, document versioning, and conflict-free multi-user editing.
Combined fast note capture, structured organization, and shared editing in one interface.
A streamlined note-taking application with Markdown support, real-time collaboration, and lightweight organization tools.
Kept email data private while providing fast AI-assisted email operations.
An AI-first email client built around a local language model for privacy, data sovereignty, and fast on-device operations.
Supported first-class functions, closures, and interactive program evaluation.
Built a complete interpreter in Go for a custom language, including a lexer, parser, evaluator, first-class functions, closures, and a REPL.
Implemented process creation, coordination, and job control through Unix system calls.
Developed a Unix shell in C with piping, redirection, background processes, process management, and custom built-in commands.
Reduced allocation overhead and fragmentation while approaching the reference implementation.
Implemented custom malloc and free primitives in C, optimizing allocation speed, fragmentation, and memory usage to approach the published reference implementation.
Processed approximately 595,000 messages per second with 1.68 microseconds of latency.
Built a zero-copy Nasdaq ITCH 5.0 binary decoder processing approximately 595,000 messages per second at 1.68 microseconds per message. Added profiling, validation, signal handling, and CSV export.