1+ year of commercial experience · open to projects

Max Saiets

Python & AI engineer. I build production systems that work with real-world data: computer vision on phone video, LLM agents, 3D models generated from map data, and the automation that runs an online store. From prototype to deploy, on my own.

Below: 10 projects, what each one solves, how it works step by step and the numbers it delivered. The three main ones run live right on the page.

10
projects shipped or in R&D
1+
year of commercial experience
1
own product live with payments
4
areas: vision, LLM, 3D, automation

About

What I do

  • Computer visionpose and video analysis
  • LLM agentstools, retrieval, evaluation
  • Geo and 3Dmaps into printable models
  • Automationbots, queues, feeds, CI/CD

Stack

  • Backend

    PythonFastAPIaiogramCelerypandas

  • Vision & ML

    PyTorchONNX RuntimeMediaPipeRTMPoseOpenCV

  • LLM

    OpenAIGeminiClaudeChromaDB

  • 3D & geo

    trimeshosmnxOpen3DBlenderThree.js

  • Frontend

    TypeScriptReactNext.js

  • Infrastructure

    PostgreSQLRedisDockerGitHub Actions

01 · 3D and geodata

Maps you can hold

OpenStreetMap and elevation data turned into clean, printable geometry: 3D city models for sale, digital twins, point clouds for simulation.

Geo & 3D · own product · live

monadruk

The monadruk generator turns maps and terrain into a model for 3D printing. Here the same approach runs on the real Matterhorn: heights from open satellite data, 4 × 4 km, sliced into 0.2 mm layers. Scroll and it prints.

How it works →
0–
80 × 80mm, the model
0.2mm per layer
DEM → meshheights → model → layers
Drag to turn the mountain

Real prints

What customers get: printed maps of real places

Also in this area

R&D · drones and synthetic data

Real2sim

Drone video becomes a photorealistic 3D copy of the place, where drones can be trained.

  • 597k Gaussians from 88 photos for $0.75 of cloud time
  • PX4 flight simulator over the digital twin
  • Automatically labelled data for training detectors
Open the case →

02 · Computer vision

Human motion from one video

Pose, 3D body and measurements from an ordinary phone video, with an honest error bar on every number.

Computer vision · 2026 · R&D

jumpsight

One ordinary camera, one jump. On the left is what the system measures, on the right the original. Scroll to move through the clip; every number is read from the system's own run on this video.

0.00revolutions in the air
–phase
–left knee, °
–right knee, °
±0.54rotation error, measured on 1,705 jumps (MMFS). 959 commits.
Axel91%
Toeloop89%
Loop80%
Salchow77%
Lutz / Flip76.5%
Jump-type accuracy on 1,688 predictions (FS-Jump3D). Lutz and Flip differ by the blade edge, which 30 fps video often cannot show.
How it works →
jumpsightoriginal
0.00 s
Clip: FS-Jump3D research dataset (CC BY-NC-SA 4.0), triple Lutz. From this video alone the system measured 2.20 ± 0.54 revolutions in the air and "Lutz or Flip"; motion capture says 2.39.

Also in this area

3D animation · R&D

Lodyk

A coach films an exercise on a phone. The system lifts the movement out of the video and a 3D skater repeats it, so the skating app can teach the exercise without showing the coach's footage or voice.

  • 21 lessons turned into 3D demos for the JumpSight app
  • A demo is published only if it passes accuracy checks: torso within 6 px, visible arms within 12
  • The coach's words voiced in English and Ukrainian
Open the case →

03 · Automation and AI

Systems that work every day

Telegram bots, job queues, marketplace feeds, LLM product copy and CRM integrations running for a real business.

Automation · in-house · 2025–26

One upload, six jobs done

At a clothing manufacturer, managers send a product video to a Telegram bot. Queues take it from there. Scroll to follow one upload.

How it works →
Manager → bot
MP4dress_5521.mp4
85commits, main bot
git pushis the deploy
YouTube✓
original upload
Telegram✓
video with overlay, to the team
JPG✓
600×900
photos resized
Gemini✓
description, checked
Rozetka · Prom · Epicentr✓
<offer id="5521"> <price>…</price>
marketplace feeds
SalesDrive✓
SKU 5521syncprice✓
CRM prices

Also in this area

Product · e-commerce

Telegram shop

An online shop inside Telegram, with no separate website.

  • Catalogue, cart and checkout right in the chat
  • Delivery by Nova Poshta or Ukrposhta, the buyer chooses
  • Every order lands in the team chat and the CRM at once
Open the case →
Telegram shop
Client project · e-commerce

fish-sync

Stock from a shop's warehouse program turns into product cards on its website, Rozetka and Facebook by itself.

  • 53 models found among 85 items, 81 variants grouped automatically
  • About 22 AI descriptions a day, each approved by a person in Telegram
  • One warehouse, three sales channels
Open the case →
fish-sync
Desktop tool · in use

Stocktake counter

A small window on top of the warehouse program that counts every scanned item during a stocktake.

  • You see how many pieces were scanned, not only the total in hryvnias
  • Tells a barcode scanner from someone typing
  • Exports to CSV and Excel
Open the case →
Stocktake counter
AI pipeline · RAG

AI Video Studio

From a library of books the system makes a narrated, illustrated video on its own.

  • 8 stages, from script to upload, with no manual step
  • The script is written only from passages found in the books
  • A quality check before any money is spent on generation
Open the case →
AI Video Studio
Personal tool · LLM

JobHunter

A personal trainer for getting hired: voice mock interviews, a study course with an AI mentor, and a job feed that picks the vacancies worth the effort.

  • Voice mock interviews from 374 real questions, with 4 interviewer styles from friendly to strict
  • Each answer is scored against a checklist; filler words and pace are measured too
  • Weak topics come back for review, and a readiness score shows progress per topic
Open the case →
JobHunter

Working together

What I can build for you

  1. 01

    E-commerce automation

    Marketplace feeds for Rozetka, Prom, Epicentr and Horoshop, stock sync from UkrSklad, product descriptions written by AI and checked before publishing.

    Examples:Commerce automationfish-syncStocktake counter

  2. 02

    Telegram bots and shops

    Bots with queues that never block the chat, WebApp shops with a cart, delivery and CRM integrations.

    Examples:Telegram shopCommerce automationJobHunter

  3. 03

    Computer vision and video AI

    Pose and motion from ordinary phone video, measured and honest about uncertainty; video generated end to end.

    Examples:jumpsightLodykAI Video Studio

  4. 04

    3D, maps and data pipelines

    Map and elevation data turned into printable models and digital twins; geometry that survives real data.

    Examples:monadrukReal2sim

How I work

  1. 01

    Understand the data

    I start from your real files, chats and edge cases, not from an ideal spec.

  2. 02

    Prototype on real input

    A working slice within days, so we argue about results instead of slides.

  3. 03

    Measure and write down limits

    Every system gets a check it has to pass, and the cases it cannot handle are documented.

  4. 04

    Ship and keep it running

    Deployment, CI, logs and alerts. I use AI coding agents for speed and keep tests and benchmarks as the judge.

Experience

  1. 2025 – now

    Automation engineer, in-house

    About a year building the automation behind a manufacturer's online sales: Telegram bots, job queues, marketplace feeds, AI-written product copy, sales analytics and CI/CD.

  2. 2026 – now

    monadruk.com, own product

    Printable 3D models of real places, from map data to payment. Live.

  3. 2026

    Computer vision and 3D motion R&D

    jumpsight, Lodyk and Real2sim: pose estimation, motion solving and synthetic data.

  4. 2025 – 2026

    Tools for small businesses

    Stock sync and marketplace feeds, a stocktake counter for a warehouse, a Telegram shop.

  5. 2023 – 2025

    Web development

    Full-stack projects in React, TypeScript and Node: shops, journals, kanban boards.

CV →

Contact

Tell me about your task

Earth imagery: NASA GIBS (VIIRS on NOAA-20, true colour, previous day), NASA Visible Earth (Blue Marble by month, Black Marble). ISS position: wheretheiss.at. Relief: Solar System Scope, CC BY 4.0, via the three.js examples.