I design and build production AI products end to end, from the model pipeline to the database to the last screen. After 18 years in enterprise IT, ending as Senior Manager of the Analytics Center of Excellence at Flex, I've spent the last three years shipping AI systems for law firms, a radiology practice and thousands of developers.
Open to AI architecture, technical leadership and founding-team roles
18+years in enterprise IT and analytics
100+patents filed with my drafting system
2,100radiology reports in Oct 2026
600+organic users on Revibe
Selected work
Systems in production
Patent Drafting AI
For IP Dome, Chennai · 2025–present · Architect & builder
Turns a university's invention disclosure form into a filing-ready patent draft, following the firm's own two-stage workflow.
Small model, structured promptsRuns on Gemini 2.5 Flash. Section-specific prompts and structured inputs give a small, fast model the accuracy a legal document needs, at a fraction of a large model's cost.
Swappable search layerPrior-art search moved from a USPTO service to search APIs to AI-based search over Google Patents, with no change to the drafting workflow.
Built and licensed independentlyDesigned and shipped in 3–4 months, licensed to IP Dome on a usage-based model, and now being prepared for other IP firms.
Impact
A search report that took a patent agent 3 days now takes under 30 minutes. 100+ patents filed.
Gemini 2.5 FlashPython · FastAPIReactPostgreSQL
AI Radiology Reporting
For a radiologist in Australia · 2026–present · Architect & lead developer
Radiologists dictate reports that usually go to offshore transcriptionists and come back 4–5 hours later. This app turns dictation into a finished draft in the practice's own templates.
Rescued and rebuiltTook over after a previous developer left before go-live. Rebuilt the product from scratch and reached the earlier feature set in 2 weeks, against a 2-month estimate.
Template-driven formattingThe LLM layer strips speech artefacts and fits each dictation to the practice's templates (shoulder X-ray, chest X-ray, ultrasound and more), drawing on reference reports when needed. Live transcription supports quick corrections.
Privacy by deletionZero-retention LLM processing, no patient identifiers in transcription, IP-restricted Cloudflare R2 storage, and all recordings and files deleted once a report is finalised. Migration to Australian hosting is underway with the practice's IT team.
Impact
Turnaround cut from 4–5 hours to under a minute. Live since September 2026, with about 2,100 reports in October.
Teaches software architecture from real codebases. Point it at a repository and it produces architecture maps, execution flows, module breakdowns and guided code walkthroughs.
Repo or ZIP (≤200 MB)→Object storage→Long-context analysis→Mermaid diagrams + docs→App · MCP · Chrome
Long-context model with a fallbackGemini Flash's large context window handles whole-repo analysis, with Claude Haiku as a backup where diagram generation needs it.
Code pulled in on demandRepositories stay in object storage and are fetched to the server only when a user opens a walkthrough, which takes a couple of seconds and keeps compute lean.
Built for AI workflowsMCP servers for npm and Python, a Claude Code skill and a Chrome extension export architecture context straight to coding agents.
Impact
Demand tested on Twitter before building. 600+ users, all organic, with steady sign-ups from Reddit.
lynclearn.com · 2023–24 · Founder & architect · Pre-incubated at Atal Incubation Centre, Anna University
Generates personalised upskilling courses built on what the learner already knows. A Python developer learning Java gets Java mapped onto familiar Python ideas; a Java developer heading into AI covers the Python they need first.
Long-form output on small context windowsBuilt on GPT-3.5 and GPT-4. Each topic was generated with the list of topics already covered, which removed repetition and made courses of up to 10 hours possible.
Synchronised audio-visual deliveryA custom pipeline aligned Google Text-to-Speech narration with Reveal.js slides, before AI video and image generation were practical.
Accuracy by scopeContent focused on concepts, where LLMs are reliable, rather than problem-specific code, where they hallucinate more.
Status
Now being folded into Revibe as its learning layer.
GPT-3.5 · GPT-4Google TTSReveal.js
How I build
Architecture principles
Right-size the model
A small model with structured inputs often beats a large one on cost and speed. I start small and add capability only where the output needs it.
Keep providers swappable
Search APIs and models change fast. Each external service sits behind its own layer so it can be replaced without touching the workflow.
Fit the real workflow
Every system follows how the experts already work, with review points where a human signs off, so adoption doesn't depend on retraining people.
Delete what you don't need
Sensitive data is stored only as long as the task requires, behind restricted access, and processed by providers that don't retain it.
Career
Experience
2023 – present
Founder, Tuple Ventures
Designing, building and licensing AI products for legal, healthcare and developer education.
Jun 2016 – Jun 2023
Flex · Senior Manager, Analytics Center of Excellence
Led the enterprise analytics function at a global electronics manufacturer.
Jan 2005 – Jun 2016
TCS · Motorola Mobility relationship
Data analytics and data engineering roles across Chennai, the US (5 years) and Singapore (1 year).
TableauAlteryxSnowflakeInformaticaPythonOracleSQL
Let's talk
I'm open to AI architecture, technical leadership and founding-team roles. If you're building something where the model is only half the problem, I'd like to hear about it.