Madhur Goel

Build an HFT platform for a 100+ trader desk where every microsecond of OS scheduling jitter was a liability.

• Bypassed the Linux kernel entirely with a user-space network driver to eliminate OS scheduling jitter from the critical path.
• Replaced OS-level IPC with lock-free shared memory queues. Cache-line alignment of all hot data structures to minimize L1/L2 cache misses was the most time-intensive optimization across the project.
• High-speed lock-free logger writing to shared memory, with a dedicated thread handling disk synchronization, keeping the trading path clear of I/O blocking.
• Cleared NSE black-box certification and algo-approval. Built and maintained as a solo engineer throughout.

Run real-time AI inference for 40 students simultaneously inside a browser, on a teacher's home network, without degrading audio or video.

• Architected a 6-region AWS platform with state in European data centers. Students connect to regional servers; teachers connect to the Indian region, minimizing round-trip latency on both ends.
• Separated video, audio, and data into three independent streams. Video degrades under low bandwidth while audio and data continue uninterrupted at 400kbps student-side.
• Deployed MobileNet V3 via TensorFlow.js, throttled to 1fps to share the browser's compute budget with concurrent audio decode, video decode, and real-time whiteboard.
• Distracted students surfaced automatically to the teacher's feed, reducing monitoring load across all 40 sessions.

Build an ML pipeline flexible enough to train and deploy across entirely different domains without rebuilding from scratch each time.

• Built the model-independent CI/CD training and deployment pipeline before any production use case, enabling rapid cross-domain experimentation.
• First deployment: automated visual defect detection for a bottle manufacturing plant. Second: real-time image personalization for consumer brands using custom StyleGAN architectures.
• End-to-end data pipeline covering collection, PII removal, and tagging across public and proprietary sources. Data preparation was consistently the most time-intensive phase of each new deployment.

Production AI systems and intelligent infrastructure: agentic pipelines, hybrid RAG, and statistically-driven payment routing.

• Hybrid RAG pipeline combining dense vector search with knowledge graph traversal for joint retrieval over text and image corpora.
• Agentic portfolio risk analysis system across equity, equity derivatives, and currency derivatives, replacing manual analysis workflows.
• Payment orchestration engine handling 1,500+ transactions/minute on Indian UPI rails. Gateway routing uses a parameterized 2σ deviation threshold, outperforming regression and decision tree baselines in live testing.
• Building DSPy-based modular AI agents with built-in prompt optimization, portable across models and inference backends by design.