Things I've taken all the way to shipped.

The problem

Specialty insurance submissions arrive as a pile of documents. Brokers, underwriters and carriers were re-keying the same data by hand — ~45 minutes per submission, across half a dozen forms.

What I built

• Built a Claude-based document extraction & intake engine that auto-populates ~70% of submission fields from uploaded documents and third-party property / hazard sources (OCR, HazardHub).
• Engineered an automated form generation & e-signature pipeline (Anvil) that produces 6+ ACORD forms (125/126/140) and carrier-specific supplementals in seconds instead of hours.
• Added carrier appetite matching that routes each submission to the markets actually likely to write it.

The problem

Tax prep and company registration meant slow, manual data entry from payslips, T-10 forms and national IDs — with no room for the two other national languages people actually use.

What I built

• Developed an AI multi-agent system that acts as an intelligent tax consultant.
• Built a tax tool that automatically extracts and structures data from payslips and T-10 forms, eliminating manual entry.
• Shipped a multilingual NIC OCR system that reads Sinhala / Tamil IDs and auto-populates director details in a business-registration platform via a custom API and frontend.

The problem

A dealership wanted to meet every customer instantly — onboarding, negotiating, booking test drives — without dropping the ones that came in after hours.

What I built

• Designed and built a full-fledged AI-powered sales agent handling onboarding, price negotiation, test-drive appointments and order placement.
• Developed a robust ERP system to streamline operations and manage customer interactions.
• Used ML techniques to improve agent efficiency and automate the sales workflow end-to-end.

The problem

Before the agents, there were pixels: surveillance, robotics, manufacturing and e-commerce problems that needed models fast and accurate enough to actually deploy.

What I built

• Built an end-to-end YOLO object-detection pipeline optimized to run at 20 FPS for surveillance and autonomous navigation.
• Engineered a stereo + monocular computer-vision dimension-measuring machine, improving measurement precision by 20%.
• Developed a cGAN (pix2pix) image-to-image translation model for editing, style transfer and colorization — a 10% lift in generated realism.
• Created a real-time monocular depth-estimation model with a 25% reduction in compute overhead for resource-constrained deployment.