Hyperion Future Tech Ventures Pvt. Ltd. (Gudz)
Software engineering on production AI-powered product systems: an agentic storefront builder on top of a live ERP, and the real-time media and inference services behind a warehouse monitoring platform. Most of my time goes to backend architecture, the AI tool harness and the reliability work that keeps both honest.
Core stack
Featured
Hyperion — AI Storefront Builder
An AI-assisted no-code ecommerce storefront builder. Merchants describe what they want in plain language and the system creates and edits multi-tenant storefronts on top of the catalogue, inventory and orders already held in the ERP — built with Next.js, TypeScript, the Vercel AI SDK and the Puck visual page editor.
The engineering idea is that the model never touches merchant data. It reads, and it writes proposals. Every change moves through proposal, validation, merchant approval and verification before anything is applied, with optimistic locking and stale-read protection so a proposal built against an older version of a shop is refused rather than silently overwriting a newer edit.
Engineering focus
Technologies
Architecture
- 25
- 11
- 14
- 0
Proof
- 1,628
- 45
- 22
- 30
Inside the build
Selected engineering work
AI harness
Typed tool definitions over the Vercel AI SDK with per-turn step budgets, a staleness guard that refuses proposals built from an outdated read, and a durable proposal lifecycle kept out of the model's context. Fixed two accuracy failures: the model answering questions instead of acting, and claiming work it had not done.
Theme architecture
First-class light, dark and system theming delivered additively, driven by one attribute on the storefront root and resolved entirely through CSS custom properties — no per-component branching, and existing shops render unchanged. The system preference is handled in CSS rather than JavaScript so it stays valid under server rendering.
SEO & structured data
Sitemap coverage, canonical handling for paginated and sorted views, and Organization, WebSite, BreadcrumbList and ItemList JSON-LD emitted only from stored values. A content-quality gate rejects keyword stuffing, unverifiable claims and thin copy, so the generator cannot invent product facts to fill a field.
Validation & regression testing
Schema validation that rejects unknown keys on write while staying lenient on read, plus a colour-contrast check that blocks unreadable palettes without punishing shops for problems they already had. Verification scripts fail loudly on missing fixtures rather than exiting green with nothing executed.
Also
Real-time warehouse monitoring
A real-time monitoring platform where mobile video is streamed to backend services and consumed three ways at once: object detection, vision-language analysis of on-screen text and unfamiliar objects, and live 3D reconstruction of the space being walked.
I moved the media layer off a peer-to-peer mesh onto an SFU over WHIP and WHEP, so the phone uploads one copy regardless of how many viewers are watching, and rewrote reconstruction from a batch job into a streaming one.
Time to first geometry
6.6s → 0.3s
- Batch to streaming reconstruction — frame-by-frame inference with incremental fusion, validated as identical to the batch path before it replaced it.
- Bounded work queues with backpressure, so capture that outruns inference degrades predictably instead of collapsing throughput.
- Point-cloud results streamed to the browser as deltas while the walk is still happening, rather than after it finishes.
- Multi-viewer media architecture — publish once, fan out server-side, no per-viewer encode on the device.
- A three.js point-cloud viewer with camera trajectory, playback and fly-through navigation.
Stack