Builds
Products I design and build myself
I lead products, and I still build them. Across eleven years I have designed and delivered more than 20 products end to end with engineering teams. The four below are my most recent, and I designed and wrote them myself with AI agents, from the data model to the last edge case.
Built by me
The Payment Reconciliation Engine
Matches every marketplace and website order to the money behind it, so leakage is found before the month closes, not after.
Four matching passes. Bank reference, exact amount, tolerance, and AI-assisted contextual matching for whatever remains.
Automatic payment classification. Each order is routed to cash on delivery or prepaid reconciliation from fields already in the order, with an explicit bucket for anything ambiguous.
Settlements unpacked first. Commissions, fees, taxes and payouts are separated before matching, including India's TCS and TDS rules for marketplace sellers.
Leakage detection. Fee overcharges, missing payouts, return losses and courier cash-on-delivery shortfalls.
Built by me
The Jewellery ERP
An ERP for fine jewellery, where every price moves with the metal market.
Live pricing. Net metal weight times the day's rate by karat, plus making charges, stone value and the correct GST split.
Compliance and stock. Hallmark (HUID) verification, certificates and piece-level inventory.
Workshop operations. Made-to-order jobs, karigar job work with material reconciliation, repairs, and old gold exchange at the counter.
Trade and retail. Memo and consignment stock, a dealer portal, gold savings schemes and a point of sale companion.
Built by me
Returns and Exchanges, Perfected for Shopify and Adobe Commerce
Post-purchase apps that run returns, exchanges and refunds for brands on Shopify, Adobe Commerce and other commerce platforms.
Every refund route. Refund to source, refund to store credit, or a split between the two.
Exchanges at the right moment. An exchange order is created only after the returned item reaches the warehouse, never before.
Unit-level accuracy. Return eligibility per shipment, item and unit, tracked in a quantity ledger, so partial and repeat returns stay correct.
Safe by design. Bundles, gift returns and admin-controlled rules, with idempotent webhooks so no refund is ever paid twice.
Built by me
Folio: The Platform Behind This Site
A multi-tenant portfolio platform for professionals, the one this site runs on.
Themes and pages. A library of themes and a block-based page builder.
Built for AI clients. A REST API and an MCP server, so AI assistants can read and update a portfolio directly.
Insight. Privacy-friendly analytics, a profile strength score and a job description match tool.
Built with AI
Agent systems I have built
Products are half of what I build. The other half is the machinery that lets AI agents work to my standard without me in the loop: reusable skills, quality gates and repository rules that turn expertise into a system.
Enterprise document engine
Produces board-ready documents and runs four automated quality gates before a person reads a word: structure, banned characters, repetition and layout.
Presentation engine
Learned a brand's visual language from its master deck, and now builds client decks in that language from content alone.
Design system for agents
Tokens, components, page recipes and agent instructions, so every agent-built website stays on brand instead of reinventing the design.
Competitive intelligence
A skill that monitors rivals across every product module and turns teardowns and gap analyses into a standing view of the market.
Editorial engine
Encodes my standard for long-form writing, from structure to sourcing, so every essay meets the same bar.
Agent constitutions
CLAUDE.md and AGENTS.md rules in every repository that fix language, testing, style and review for every agent that touches the code.
AI expertise
Ten stages of AI expertise. I lead from the tenth.
I measure AI expertise by the errors a person can catch, not by the answers they can get. This is the ten-stage model from my essay The Ten Stages of AI Expertise.
I work at the tenth stage, and I still do the work of every stage below it myself.
The ten stages
Tourist
Asks and accepts the answer, catching only obvious nonsense. The starting point for everyone.
Operator
Uses AI daily with reusable prompt patterns, catching problems of format and tone.
Context engineer
Supplies the right source material, catching answers that lacked what they needed. I give every agent the raw evidence, not a summary of it.
Verifier
Checks claims with methods independent of the model. Every agent output I accept is checked against the original source.
Workflow designer
Breaks work into steps with checkpoints. My agent workflows define the evidence each step must show before the next one starts.
Evaluator
Builds test sets and measures error rates. I wrote the case for proving failure rates before launch in AI Product Readiness Is a Rate Claim.
Builder
Ships AI features with retrieval, tools and guardrails. I build and ship with Claude Code, Cursor and MCP every day, and lead AI product information management in production.
Economist
Prices cost, latency and reliability. I judge AI on completed work, as argued in The Economics of AI Begin After the Answer.
Systems owner
Runs production AI end to end, including incidents, safety and governance. I own AI products in production and the escalations that come with them.
Multiplier
Shapes how a team or organisation works with AI. I won the company's Generative AI Crackathon, wrote the playbook on leading AI-enabled teams, and build the reusable skills and workflows my teams use.
Depth
AI, inside out
Knowing how to use AI tools is the entry point. Knowing how every layer underneath them works is what lets me decide where AI belongs, how far to trust it, and what it should cost.
| Layer | What I know | Where I use it |
|---|---|---|
| Models | Frontier and open models, their strengths, limits and failure modes | Choosing hosted or local models per task, including open models through Ollama |
| Context engineering | What a model must see to answer correctly | Every agent gets the raw source, not a summary of it |
| Retrieval | Search, embeddings and grounding in the business's own records | Research and analysis over mail, documents, meeting notes and CRM data |
| Tools and MCP | Scoped, safe access from models to real systems | Folio exposes an MCP server, so AI assistants can update a portfolio directly |
| Agents and orchestration | Multi-step and parallel agents with checkpoints between steps | Several coding agents in parallel, each on its own repository and task |
| Evaluation | Test sets, error rates and failure modes measured before launch | Proving failure rates before an AI product ships |
| Economics | Cost, latency and reliability per completed task | Judging AI on finished work, not on answers |
| Safety and governance | Guardrails, data boundaries, prompt injection risk and audit trails | Quality gates in every agent skill, and repository rules for every agent |
| Vision and multimodal | Image understanding across products and catalogues | Catalogue enrichment, AI skin analysis and virtual try-on |
Stack
The stack I build with
- TypeScript
- JavaScript
- Node.js
- Next.js
- React
- SQL
- MySQL
- PostgreSQL
- Prisma
- MongoDB
- Redis
- REST APIs
- GraphQL
- Webhooks
- EDI
- Shopify Admin API
- Claude
- Claude Code
- Cursor
- MCP
- Postman
- JMeter
- SoapUI
- Ollama
Work with me
Let's build what comes next.
I work with operators, product leaders and founders on commerce platforms, marketplaces and building with AI. A short conversation is the fastest way to find out whether I can help.