Ishaan Goyal

I helped build the ML and data systems behind interactive music products at a startup valued at $50M, working alongside Spotify's former Head of Innovation and turning customer feedback into product improvements.

I fine-tuned BS-RoFormer for stem separation, optimized it with sparse attention and weight quantization, and deployed it for on-device inference on iOS. Alongside the model work, I built an internal ingestion pipeline for 1M+ songs.

I liked moving between model behavior, device constraints, and product ideas with a team that wanted people to play with music.

I worked on ConstraintFlow, a compiler that turns neural-network verification specifications into tensor computations. My focus was the compilation path: refactoring IR passes, adding caching and JIT compilation, and profiling more than 20 passes to find bottlenecks. I also coauthored the team’s compiler and runtime preprint and designed a roadmap for CPU–GPU synchronization and multi-GPU scaling.

I built an inference runtime for speculative and speculative speculative decoding with fused GPU kernels, CUDA graph replay, and execution overlapped across two GPUs. With Qwen2.5-7B in FP8 and a trained EAGLE-3 drafter, it reached 697 decode tokens/s on one request on two B300s, about 2× my tuned autoregressive baseline, and ran a single request to 128,000 tokens. Along the way I built an exact-match test suite, tracked down a KV-cache stride bug, and found new matmul schedules worth 17–21% end to end.

I built a decoder-only Transformer in JAX, Flax, and Optax for arithmetic reasoning, then extended it with top-1 mixture-of-experts routing. From there, I wrote a fused Pallas projection kernel to reduce activation memory traffic and compared it with a ragged-dot baseline. The project follows the same workload from model design down to accelerator execution.

I profiled GPT-2 inference on an NVIDIA A40 to understand where GPU time was going, then focused on matrix multiplication. I explored shared-memory tiling, launch configuration, vectorized loads, and a WMMA path, and added KV caching for autoregressive generation. I compared kernel behavior with the full inference path to understand how local improvements carried through to the application.

I helped build Ren on VS Code and ship it to alpha users. My work included codebase indexing and retrieval, interactive agent workflows, and agent orchestration. I also built views for agent activity, tasks, and file changes, alongside patch review and an execution timeline. The goal was to give developers enough context and visibility to stay involved as coding agents worked.

I built an event-driven video pipeline that sends multipart uploads directly to S3 and turns them into adaptive HLS streams. RabbitMQ coordinates Go workers for FFmpeg transcoding and captions, while GraphQL subscriptions keep the interface updated as jobs run. The media processing runs in the background, separate from the interactive upload and playback flow.