Hard problems, hard solutions.

A selection of the hardest problems I've solved : and what they took to get there.

Client names omitted where NDAs apply. Details available under NDA.

Ultralow latency High-Frequency trading platform

Context: Built from scratch for a third-generation proprietary trading desk (100+ traders) in India.

Need: A platform that could pass NSE black-box certification : algos, Orderbook builder, order manager, and risk audited to the packet level.

Work: End-to-end low-latency stack; market connectors and risk engine for NSE; led a team of 3 through design, implementation, testing, and certification; full documentation for algo approval.

Outcome: Certified and algo-approved by NSE. 2 μs end-to-end latency (NIC → NSE gateway ack) at the 99.2nd percentile.

C/C++ · low-latency networking · NSE compliance · kernel Bypass · Cache hits

Enterprise-grade generative AI (image personalization)

Context: Co-founder/CTO at a San Francisco computer vision startup (tanget.ai) incubated in Alchemist Accelerator.

Need: Real-time image personalization for a major fashion/retail brand : dynamic catalog imagery to improve campaign efficiency.

Work: Custom StyleGAN on proprietary imagery; real-time generation at scale; first enterprise client during the program; seed raise with technical differentiation.

Outcome: Among the first enterprise-grade consumer-facing GAN deployments; measurable campaign efficiency gains.

StyleGAN · CUDA · PyTorch · serving infrastructure

Video confrencing with Real-time AI for developing-world networks

Context: Co-founder/CTO at a video confrencing SaaS company (2020) for EdTech during COVID.

Need: Live Video at 400 kbps, 40-student sessions with in-browser AI, 99.9% SLA.

Work: 6-region AWS architecture; full in-browser AI pipeline (attention, behavioral signals); WebRTC tuned for constrained networks; SLA design for voice and whiteboard.

Outcome: Pan-India EdTech deployment; zero AI-layer crashes over long sessions.

WebRTC · multi-region AWS · TensorFlow.js · SLA design

Payment orchestration on Indian UPI rails

Context: Fractional CTO at a payment technology company (2025–present).

Need: Serverless Intelligent routing at high volume : gateway choice from live conditions.

Work: Orchestration engine at 1,500+ transactions/minute on UPI; routing from failure rates, banking preferences, and capacity; integrations and monitoring.

Outcome: Automated routing replacing manual decisions; zero downtime at peak load.

High throughput · UPI · Python · real-time monitoring

Autonomous risk agents & Prompt optimization strategies

Context: Active consulting (2025–present), financial/fintech.

Need: Reliable retrieval on structured and unstructured data; automated risk analysis without constant human intervention.

Work: Hybrid retrieval (vector + knowledge graphs); LLM-as-Judge evaluation; autonomous portfolio risk agent across equities and derivatives; DSPy for prompt optimization.

Outcome: Replaced heavy manual review with monitoring and alerts; retrieval tuned via automated evaluation with human validation.

RAG · KGs · agents · DSPy · Python

Have a problem that looks like this?

I've probably built something close to it. Let's talk about yours.

Start a conversation →