Lev Skorokhodov

• Telegram bot read photos of handwritten invoices with Claude Sonnet on AWS Bedrock, matched lines to the live catalog and queued them for approval. The public demo bot runs the same flow on GPT-4o Vision.
• The bot handled 10 to 30 invoices a day, peak 60.
• The client bought out the code. The AI part is rewritten from scratch as a public repo: extraction, catalog matching, approval queue, eval on synthetic invoices.
• 0 wrong catalog matches on 124 lines of 20 synthetic invoices: an unreadable name goes to review instead of a guess. $0.003 to $0.026 per invoice across 3 models.
• 9 modules: sales, AI inbox, purchases, trips, FIFO warehouse, deliveries, price list, settlements, catalogs.
• Built solo. MVP in 5 days from scratch.
• Reverse proxy in Moscow kept the offshore server reachable for staff in Russia.

React 18, Vite, Node.js, Express, PostgreSQL, Claude on AWS Bedrock, OpenAI API (demo), Telegram Bot API.

• Agent referral links, 90 day first touch attribution, UTM and click ID tracking, CPL, CAC and ROAS.
• Every agent has a cabinet with a personal link and their own clicks and leads.
• Lead dedupe by phone or Telegram. Apartment statuses sync back to the site.
• The sales site feeds it: apartment picker over 279 units with floor plans and a PDF plan for every unit.

JavaScript, Node.js, Fastify, SQLite, nginx.

• Reads PDF layers instead of guessing by color: 39k objects per floor down to 1.9k.
• Areas match the official schedule within 0.05 m².
• 5 Sonnet agents in parallel close 25 floors in about 2 minutes.
• Manual edits in Figma are diffed by object ID and become batch rules for the next run.
• Every write passes a 3 second Playwright render check. Snapshot at every stage.

Python, PyMuPDF, shapely, Figma MCP, Playwright, Claude Code subagents.

• One React app runs as a Telegram Mini App and as a website. Mini App requests are checked by HMAC of Telegram initData.
• Bot wizard: photos, nested model picker, condition, price, then a 1:1 preview of the channel post.
• 340 tests run offline in under 20 seconds.
• Built by an agent loop: Sonnet subagents write routine parts, Opus the complex logic, acceptance by diff, pytest and curl on staging.

Python 3.12, FastAPI, aiogram 3, SQLAlchemy 2, SQLite, Alembic, React 18, Vite, TypeScript, Tailwind.

• 6 stages: Whisper word timestamps, typo check, GPT-4o scene picking, punctuation, FFmpeg render.
• Content hash cache: a repeated upload costs nothing.
• Transcription is half the cost: $0.14 of $0.27, measured on 10 episodes.
• Review UI with live logs over SSE, trim handles, banner editor, render queue, Telegram publishing.

Python, FastAPI, SSE, OpenAI API, FFmpeg, ASS subtitles.

• 6 campaigns, 103 planned shots, 244 generated frames, 47 clips.
• Technics SB-MX200 spec film: 27 seconds cut from 38 generated clips. The choice between Seedance and Kling is settled by measured motion, not taste.
• Storyboard with a job bus: a button on a shot wakes Claude, which picks up the job. After 16 operator rules, credit waste fell from about a third to zero.

Higgsfield CLI, Nano Banana Pro, Seedream 5, Seedance 2.0 and 2.5, Kling 3.0, Palmier over MCP, Claude Code skills, Node.js.

• 25 tools behind one MCP server. The web agent, Claude Desktop and Claude Code call the same endpoint.
• Two phase agent: discovery on 11 tools, analysis on 17.
• 75 tests, mypy strict in CI. Built in 4 days.

Python, FastMCP, OpenAI Responses API, FastAPI, SQLite, pymatgen.