Hi, I'm Sourik

Founding engineer for backend, ML infra, applied ML direction, and product. Built Flux 1.1 training/inference infra across Modal and Datacrunch, latent interpolation upscaling, bulk moodboard templating, and Foosh v2 async multi-modal workflow engine powering 2M+ image generations.

Built a 6-phase memory pipeline with Gemini structured extraction, Pinecone hybrid BM25+dense retrieval, temporal decay, contradiction detection, and narrative summaries. Achieved 95% extraction accuracy, cut contradiction checks by 95%, and lifted unified retrieval from 62% to 81%.

Reimplemented an MLSys 2026 paper with a Rust+Python hybrid using DP range partitioning and streaming rANS entropy coding. Reached within 0.3-2.9% of Shannon's limit and beat zlib on 20/21 ResNet-18 layers with 1.55x overall compression.

Built zero-training open-vocabulary logo detection using Gemma 4 26B-A4B MoE through MLX on Apple Silicon, with four-factor QoE scoring, auto-cropped proof images, JSON/Markdown reports, and a self-contained dashboard with heatmaps.

Built a local AI radio station where ACE-Step 1.5 generates music via MLX, Kokoro TTS voices an AI DJ, and a look-ahead scheduler hides generation latency for gapless playback with zero buffer underruns.

Built sub-50ms Amazon retail ingress and anomaly alerts, migrated Razorpay tier-1 payment routing APIs to Go microservices, improved Tekion availability from 99.0% to 99.99% and latency by 90%, and shipped finance tooling for 400+ managed creators.