Kasatov AI & Data
DE Hire me

Part designation:

Sergey
Kasatov

17 yrs automotive engineering → data & AI automation

Focus: AI & Automation · Data Analytics · Cologne, DE

I design AI agents that a business can trust and I test where they break.

17 years of automotive engineering on Ford, GM, and Stellantis programs, retrained into data. Today I design and operate a multi-agent system that runs 19 workflows unattended, backed by the 8D discipline of a quality engineer.

Explore my work Get in touch
GitHub LinkedIn Tableau Email
Released for hireavailable now · DE
Fig. 1 · SPC run chart wk 1–24
UCL LCL
Issue time-to-close: ~4 wk → ~3 wk · Stellantis program
17yrs
automotive engineering: development, quality, launch
7
vehicle launches supported with plants and suppliers
€250k
largest documented cost saving per initiative
13/ 3
portfolio projects / deployed online

Section A-A · Why this profile

An engineer who automates his own work

Most analysts learn the domain from dashboards. I spent 17 years inside it: plants, suppliers, launch ramps. Then I learned the data stack, and built the agents that now run the routine part of my own work.

Manufacturing domain knowledge

I have stood at the line where the data is produced: plants, suppliers, launch ramps. I know what the numbers mean physically.

Structured problem solving

8D, SPC, PPAP/APQP: a root-cause discipline that transfers directly to analytics. Hypothesis, evidence, corrective action.

Clear communication

Years of aligning OEMs, plants, and suppliers: I explain models and metrics to non-technical stakeholders without losing the substance.

Parts list · Featured

Work that ships and keeps running

Every metric is reproducible from the linked repositories or demonstrable online: two live Tableau dashboards and a Streamlit demo app.

Pos. 01 · Jul–Sep 2026 Live
Screenshot of the Arkon Quality Steering Cell Executive view on Tableau Public, as of 6 Sep 2026 16:12 on the plant clock: the status sentence (20 of 50 open incidents open more than three days, the 6 raised in the last hour worked inside window), four KPI cards (overdue 14 of 50 open, 28 percent; open 50 with 169 raised, 113 closed and 6 resolved; time to acknowledge 48.7 min median against a window of 15 min for P1 and 60 min for P2; time to close 144.1 min median over 113 closed), open incidents by age band stacked by priority with the part past its window in red, the six slowest acknowledgements as a multiple of each incident's own window, and the six most recent transitions with who acted
Fig. 01Executive view on Tableau Public, as of 2026-09-06 16:12 · captured 2026-09-06

Arkon Manufacturing AI

7ML models behind one Quality Steering Cell

An Industry 4.0 quality platform for a fictional heavy manufacturer, deployed on a NAS and running: seven models on real public datasets publish one twelve-field risk event, an n8n Quality Steering Cell triages it and puts a Telegram card in front of a named person, a Streamlit cockpit is the only surface that moves an incident, and a Tableau executive view generated from the same store is published on Tableau Public. Two agents sit on top of it and face opposite ways: a Langflow assistant that explains to the plant without acting, and an n8n Customer Quality Desk that answers a customer from outside through a read-only projection of eight fields.

XGBoostPyTorchn8nLangflowTableau
Executive view on Tableau Public Card in the parts list Code on GitHub
Pos. 02 · since Jun 2026
Screenshot of the Vault Mission Control board, worker grid only: 25 tiles, each with the worker name, its schedule, last run, 30-day run history and status; WARN states shown as found
Fig. 02Mission Control, the live status board · captured 2026-09-03

AI-Brain Agent System

19scheduled workflows, one owner each

A private multi-agent system that runs unattended: three LLM runtimes and a self-hosted n8n server on one Markdown rule set, two schedulers on two deliberately independent machines, and five verification gates that can fail a run. Designed, specified and operated by me; the code was written largely with AI assistance.

Claude CodeCodexn8nPythonlaunchd
Case study
Pos. 03 · Detail view Demo · wakes in ~30 s
Screenshot of the FreshMart Demand Planner Streamlit app after Generate forecast: five KPI tiles (total forecasted sales 6,779, average daily 484, peak day 2014-01-05, peak sales 768, demand level high) above a line chart of recent sales and the 14-day forecast
Fig. 03Streamlit app, forecast screen · captured 2026-09-03

FreshMart Demand Planner

138.0champion RMSE

Retail demand forecasting end to end: 10+ models compared (SARIMAX, Prophet, XGBoost, LSTM), tuned with HyperOpt, tracked in MLflow, shipped as a Streamlit planning tool.

PythonTimeSeriesMLflowStreamlit
Case study Open demo Code
Full parts list: all 13 projects
Inspection report · third-party assessment Source: LinkedIn · public
“Sergey brings a rare blend of strong technical capability, analytical thinking, and excellent communication… He is technically strong, business-minded, and an excellent communicator.”
Tom Themeles Data Science Instructor, Masterschool
More recommendations

Hiring for data, automation or quality?

Open to permanent AI consulting and enablement, automation, and data analyst positions in Germany, available now. Happy to walk through a concrete data or automation problem from your team in an interview.

sergey@kasatov.de Connect on LinkedIn
Released for hireavailable now · DE