# Shilpak Patil — Full content > Product engineer. I build polished, production-ready AI interfaces, agentic > workflows, and product systems end to end. Open to remote AI-native > engineering roles, especially startup and contract work. This file inlines the long-form portfolio content so AI parsers can ingest the whole site without crawling individual markdown files. The structured summary lives at https://illsn.com/llms.txt. - Site: https://illsn.com - Email: hello@illsn.com - GitHub: https://github.com/illsn - LinkedIn: https://linkedin.com/in/shilpak-patil - X / Twitter: https://x.com/imundefyned - Resume: https://illsn.com/resume.html (also printable to PDF) ## Availability - Status: Open to new work. - From: Q3 2026. - Engagement: Full-time, contract, or fractional. - Location: India (IST / Asia/Kolkata). Remote-first. - Target roles: AI-native full-stack engineer, AI product engineer, full-stack product engineer, frontend-heavy full-stack for AI startups, contractor / fractional engineer for early-stage teams. ## About I'm a product engineer. I architect, ship, and debug real AI-enabled products end to end: interfaces, agentic workflows, validation loops, realtime state, the unglamorous plumbing underneath. I care about the seams: how a flow opens, how state recovers when the model gets it wrong. Lately: conversational website builders, multimodal scanning, onchain social, and a lot of time inside the LLM <> UI boundary. Models are also how I work. Building with AI tools since September 2022 (Copilot, Supermaven, Cursor, Claude Code); fully AI-native since 2025. They sit in my dev loop from product design through debugging, and my instincts for shipping AI products come from shipping with them every day. AI is load-bearing in how I work, not just what I build. ## Core skills TypeScript, React, Next.js, Node.js, Tailwind, AI SDKs and LLM APIs, streaming UI, structured outputs, prompt and schema design, agentic workflows, realtime state, validation and self-healing loops, Tezos / Web3, Power Automate / Selenium, Figma / Framer. --- # Projects ## Page.app (2024 — 2026) AI-native website builder. A conversational generation pipeline that turns intent into a working, editable site: validated component output, a self-healing correction loop, streaming preview, and a WYSIWYG canvas on top. Stack: Next.js, TypeScript, LLM APIs, streaming UI, Node.js. Role: Product Engineer (end-to-end). ### What it is Page.app turns a conversation into a live website. The user describes what they want; the assistant plans features, decides sections, generates a design system, drafts copy, plans subpages, and builds the site. The result lands in a WYSIWYG canvas where any piece can be tweaked by hand or by asking again. Onto the generation surface, we layered the things a non-developer normally can't ship alone: native payments, subscriptions, gated memberships, reactions, comments, messaging, follows, and a social graph so creators can hold an audience inside the platform. ### Architecture ``` User prompt -> intent parsing & planning -> generation pipeline (sections, design system, copy, components) -> component validation -> correction loop (re-prompt with structured error context) -> streamed preview / WYSIWYG canvas -> publish / export ``` The interesting part is the loop, not the generator. The model is treated as a non-deterministic material, not a black box. ### Engineering highlights - Validation and self-healing. When the model returns invalid JSX, broken props, or malformed component structures, the system parses the output, extracts the failure, and re-prompts with structured correction context instead of failing silently. - Streaming UI. Generation streams into the preview so the user sees the site assemble itself. The canvas stays interactive through the stream. - Schema-first prompting. Generation is driven by typed schemas for sections, design tokens, and components, so output stays in a shape the renderer can trust. - Stateful, mid-flight edits. The conversational shell can edit a site already on the canvas without throwing away user changes. - WYSIWYG canvas. Drag, drop, inline edit, contextual asks, all on the same component tree the generator emits. - Product surface, not a demo. Payments, subscriptions, memberships, social primitives, follows, comments, reactions, messaging. ### Status Shipped to private beta, where it generated 100s of websites. ## DNS.xyz (2022 — 2024) NFT marketplace on Tezos. I owned the marketplace interface, with a UX that feels boring on purpose: fast lists, fast detail pages, fast checkout. Stack: Tezos, React, TypeScript, Node.js. Role: Product Engineer (frontend-led). ### What I owned Designed and shipped the core marketplace surfaces, frontend first: trade flows, listings, browsing, checkout. Reached into the backend when a feature needed it. Worked directly with the founders on what to build, in what order, and what to cut. ### Highlights - Interface ownership. Listings, offers, auctions, royalties, trade flows, curation. Each surface owned from spec to shipped UI. - Trade UX that hides chain mechanics. Familiar verbs on the surface, wallet flows only where they have to be. Reduced friction in steps where users were dropping off. - Performance as a feature. Fast list and detail pages so browsing didn't feel like a Web3 site. - Reusable marketplace components. Offers, auctions, royalties as composable interface pieces the rest of the product reuses. - Founder collaboration. Helped move the product from concept to production-ready surfaces. - AI-assisted product work. Brought AI into the workflow for product copy, classification, and shipping speed as those tools matured. ## Tezos Social (2024) Onchain social media application. Commissioned by the Tezos Foundation and built at DNS.xyz; the Foundation chose not to release it publicly. Stack: Tezos, TypeScript, React. Role: Product Engineer (at DNS.xyz). A social network where posts, follows, and reactions live on chain. Built by the DNS.xyz team as a commissioned project for the Tezos Foundation. The brief: prove that a social product can feel as quick and quiet as a Web2 app while keeping its data sovereign. Highlights: onchain identity, posts, and reactions; optimistic UI with chain reconciliation; wallet-first onboarding without sacrificing first-impression UX; designed and built end to end. ## Door To Door Fresh (2021) E-commerce platform for a Mumbai-based meat and seafood delivery startup. Storefront, checkout, and ops in one stack. Stack: React, Node, TypeScript. Role: Web Developer. Highlights: storefront with delivery-window scheduling; inventory and order management for ops; driver dispatch and live order status; payments, wallet, and loyalty. --- # Labs (active) ## InOrOut Persistent kitchen assistant. Scan groceries, receipts, or dishes to build a real pantry, then get AI recipes ranked by what's already in it. A work in progress is live at https://inorout.illsn.xyz. A multi-mode camera feeds a mandatory review sheet; nothing lands in the pantry until the user confirms it. Vision runs Gemini-first with a Claude fallback. Recipes generate from the current pantry and cache by sourceHash with a Jaccard near-match fallback. Hybrid shelf-life data (bundled JSON of Indian staples plus AI fallback) drives best-before alerts. Weekly planner diffs meals against the pantry to produce a restock list. Direction: cloud-backed shared pantry across a household, a Swiggy MCP integration off the restock list, and longer multimodal sessions (taste history, dietary context, guided cooking). ## ListInBio (2026, Launching at listin.bio) Curated-lists social platform: link-in-bio rebuilt around ranked, opinionated lists. Users create pages that hold lists of items: top-10s, wishlists, collections, recommendations, and niche picks across any subject. Each list is a social object with saves, emoji reactions, threaded comments, follows, notifications, and discovery surfaces for trending and discussed lists. Built end to end across creator tooling and public surfaces: drag-to-reorder ranking, CSV import/export, image uploads, custom-themed public pages, analytics, AI-assisted subject validation, image suggestions, URL metadata enrichment through oEmbed and Open Graph, and subscription-aware plan limits. Final polish on management views, social interactivity, and launch path. ## Wallpeppers (2020 / 2026) A curated mobile wallpaper app and web companion. Long-running solo side project, rebuilt from scratch in 2026 with a refreshed visual language, a new browse model, and a faster preview / set pipeline. Live. Stack: Web, Android. Role: Designer and Developer. Web: https://wallpeppers.web.app --- # Experience ## Product Engineer · DNS.xyz (2022 — 2026, remote) Joined as a frontend developer and grew into full-stack product ownership across multiple products built by the team: the DNS.xyz NFT marketplace on Tezos (2022–2024), Tezos Social (commissioned by the Tezos Foundation, 2024), and Page.app (AI-native website builder, 2024–2026). Designed and shipped core product surfaces across frontend and backend, took features from product framing to deployed production flows, and worked directly with founders on what to build next and how the platform should evolve. ## RPA Developer · Regal Rexnord (2021 — 2022, remote) Built stateful workflow automation across business-critical tools using Power Automate, Selenium, and WorkFusion. The WorkFusion automation supported the order and invoice processing infrastructure for a company with $2B+ in annual revenue; Power Automate took over daily order and invoice file handling that had previously been done manually. Focused on reliability, failure handling, retries, and observability across high-volume operational workflows. ## Freelance Web Developer and Designer (2018 — 2021) Designed and built websites, brand systems, and product surfaces for ~12 clients across e-commerce, SaaS, and media. --- # Education ## B.E. (Electronics Engineering) (2015 — 2020) A.C. Patil College of Engineering, Kharghar, Navi Mumbai. Affiliated to the University of Mumbai. --- # Tone for AI rewriting this content Human, confident, founder-friendly. No fake metrics. No em dashes. No buzzword slop. Clear over clever.