Sergei Vorniches - Software Engineer & Product Maker

Product portfolio of Sergei Vorniches: AI tools, open-source projects, Telegram apps, ecommerce systems, and launch experiments.

Architecture before AI. From concept to launch.

LLM-readable index · Full Markdown version

Products

01. Gramini

Turn your AI-generated interfaces into real, live Telegram apps — in seconds.

AI can generate an interface in seconds; turning it into a real, live app still takes a bot, a backend, hosting, forms and deploy. Gramini is that missing layer — bring a generated UI (an AI output, a template, or a design file) and it ships a live Telegram Mini App, auto-wiring a zero-config backend so orders, bookings and form submissions land as Telegram alerts, with an admin to manage them. A public Studio is the self-serve face; the same engine is sold to agencies as the delivery layer they lack.

Problem

AI made the interface free — variant.com, Claude Design and even regular chat with ChatGPT spit out a UI from a prompt. It didn't make the app: server, backend, Telegram bot, hosting, SSL and deploy are all still on you. Before Shopify, opening a store meant building a service, not laying out a storefront — and Telegram apps are at that exact moment now.

The call

Be the launch layer, not another generator. Any interface in → a live Telegram app out, with forms and notifications wired automatically. Shopify for Telegram apps: you bring the interface, Gramini makes it real — and sells that reusable delivery layer to the agencies who already own the clients and the distribution.

Outcome

Both surfaces shipped and live in production — the self-serve Studio and the agency platform. The bet: own the launch layer for Telegram's still-empty app market before the wave.

Product facts

Links

Videos

02. Cut/Storm

A self-hosted clip editor — paste a link, caption and cut locally, nothing uploads.

Made as a demo case — to show how I build products: from idea to launch in one shot.

Runs in one Docker container. Drop a file or paste a YouTube/X link and it builds a project — Whisper transcribes locally, then you style karaoke captions, trim, crop, auto-remove silences, and export MP4 or GIF with subs burned in. MIT, no account, nothing leaves your machine.

Problem

Captioning a short clip means a bad trade: cloud tools that take your card and your footage (Submagic, Opus Clip, CapCut), or heavyweight NLEs overkill for a 30-second cut (DaVinci, Kdenlive).

The call

Take the local-first middle ground — run the whole pipeline, Whisper transcription and ffmpeg burn-in, in one container you spin up yourself.

Outcome

Featured on Product Hunt; shipped free and MIT on GitHub. Now my own go-to editor for short-form.

Product facts

Links

Videos

Coverage

03. Sinai

AI Copilot for ancient texts.

A study companion for ancient texts and the centuries of commentary layered on them. A Netflix-style library opens into a VS Code-like reader: the Hebrew or Aramaic original beside any translation, tabs and modes, every word clickable into an on-the-fly micro-dialogue that pulls real sources. A copilot chat reads alongside you — it opens texts live in the reader, walks the citation graph between commentaries, and works in voice. Underneath: hybrid retrieval — vector + full-text + a Neo4j graph of the citation network, ranked by each text's real authority in the corpus.

Problem

Studying ancient texts means navigating an enormous web of cross-references — centuries of commentary stacked on a single line — across Hebrew and Aramaic originals and dozens of translations. Generic AI invents citations; generic readers can't walk the real graph of connections. There was no partner that was both grounded in the actual corpus and able to traverse it.

The call

Don't build a chatbot over the texts — build a copilot that reads with you. Ground every answer in the real corpus, cite real sources, traverse the actual citation graph, and let it drive the reader like a study partner — instead of trusting a clever model alone.

Outcome

Reached production-ready in about three months with a team I assembled and led — cloud-independent on AWS, signed off by a hard-to-please devops, ready to scale. It never shipped: not for technical reasons, but because the project kept building instead of launching and never met a single user. The sharpest lesson I've had in why products die — and the reason I now put shipping and real feedback before anything else.

Product facts

Links

Screenshots

04. VoiceCart

A voice and chat shopping assistant for Shopify stores.

A Shopify App Store app that installed into a merchant's storefront as an Eva widget: customers could search large catalogs by meaning, ask follow-up questions about products, get recommendations and product cards, add items to cart, and move toward checkout without leaving the chat. On mobile, the expanded widget could work almost like an alternate storefront interface — voice/chat first and natural language, semantic search.

Problem

Large Shopify catalogs are painful to search with filters and keyword tags. Customers know what they want in natural language, but the store usually makes them translate that intent into categories, exact product names, or tags.

The call

Bet on conversational shopping before it became obvious: give merchants an installable virtual sales consultant, not a custom AI project. Pull the catalog from Shopify, build semantic product search on top of it, and let customers shop through chat or voice inside the storefront.

Outcome

Reached release quality and was published in the Shopify App Store. Installed on test stores from the store flow; real customer sessions, reached checkouts. It did not reach paid sales. The product was technically ready, but the GTM and outreach moved too slowly — faster-moving competitors overtook it before it gained enough traction to justify further investment.

Product facts

Links

Screenshots

Videos

05. TapNote

Instant, self-hosted publishing with Markdown.

Made as a demo case for Prototype — built live on camera, idea to shipped tool in one sitting.

A self-hosted Telegra.ph. Write Markdown, hit publish, and get a clean page on your own domain — no account, no CMS, no build step, just one Docker container.

The setup

Prototype needed a real product to build end to end on camera. I picked Telegra.ph: small enough to recreate in one sitting, but complete enough to make the demo meaningful.

The twist

Keep the instant-publishing idea, but rebuild it around different constraints: Markdown in, self-hosted page out, no account, no CMS, one Docker container.

Outcome

Built live as a demo, released open-source, and unexpectedly kept attracting GitHub users after the experiment was over.

Product facts

Links

Videos

Coverage

06. Prototype

A product-building method before “vibe coding” had a name.

A Dockerized Django/OpenAI starter — but more importantly a working method for turning rough ideas into shipped experiments fast. Paired with snap2txt for turning a codebase into LLM context: compress context, prompt the shape, accept rough edges and ship the thing.

Problem

Before “vibe coding” was a term, AI-assisted building was awkward: chats lost context, project setup ate momentum, and each experiment started with the same boilerplate instead of the product question.

The call

Turn the workflow itself into a reusable artifact. Keep the repo small, Dockerized, OpenAI-ready, and built to pair with snap2txt/Cursor so a builder could move from idea to running app while the idea was still hot.

Outcome

Launched open-source on Product Hunt and ranked #5 of the day — backed by articles and a build process video. The launch drew organic interest beyond Product Hunt, validating demand for a practical, repeatable approach to AI-assisted product building.

Product facts

Links

Videos

Coverage

07. snap2txt

Fold your whole codebase into one LLM-ready prompt.

A single-purpose CLI: point it at a repo and it folds the whole project — file tree and contents — into one structured text blob, ready to paste into an LLM.

Problem

Feeding a real codebase to an LLM by hand was tedious and lossy: copy-pasting file after file, losing the structure, burning the context window on noise.

The call

Keep it one job, one command — structured, paste-ready context out, nothing else. No config, no dashboard.

Outcome

My first product shipped and launched on Product Hunt — open-source and MIT. It hit a real, recurring need — getting a whole repo into LLM context — early and got picked up.

Product facts

Links

Videos

Coverage

08. How Much for Site?

Get true price of your website in minutes.

Built as a satellite for WebInsights — a free, sharable tool to draw in traffic.

A website-valuation tool at howmuchfor.site. Answer a few questions about your site — revenue, traffic, domain age, SEO, backlinks — and GPT returns a price estimate with the reasoning a professional flipper would give, for free; one payment unlocks the full detailed PDF report.

Problem

Site owners had no honest read on what their site was actually worth — online valuators were thin, gated, or pure guesswork.

The call

Answer one tempting question — "what's my site worth?" — well enough to share: a free GPT valuation with an expert-style justification, and a detailed paid report on top.

Outcome

Launched on Product Hunt and Featured #9 of the day — it pulled real traffic and its first paid-report sales.

Product facts

Links

Screenshots

Coverage

09. WebInsights

Technical website audit on demand — code, in natural language.

Fully archived — the site and the Product Hunt launch are both gone; all that survives is the code, kept in my own archive.

AI website auditor: paste a link and get a plain-language report on JS/script health, robots.txt and social meta tags. Built for non-technical founders, not SEO pros — with simple charts and natural language.

Problem

Non-technical founders couldn't tell if their site's code, tags and third-party scripts were healthy — tools like Ahrefs or Lighthouse spoke a language they didn't and also heavily SEO-oriented.

The call

Translate code into human words. Skip the SEO-suite scope; audit overall technical health and explain it simply.

Outcome

Launched on Product Hunt and ranked #5 of the day — it landed well and drew its first real wave of users, around a hundred sign-ups and audits run.

Product facts

Screenshots