# Sergey Kasatov - Data Analyst Cologne, Germany. Available from September 2026. This is the machine-readable mirror of https://kasatov.de/. Same information as the website, in a format agents can parse directly. The designed version lives at the HTML pages linked below. ## TL;DR - Data Analyst with 17 years of prior automotive engineering on Ford, GM, Opel and Stellantis programs: product development, supplier quality, and launch management. - Retrained into data analytics (Masterschool, Data Analytics Program AZAV, Advanced Data Science and AI). Python, SQL, Tableau, machine learning, with a preference for interpretable models. - Eleven completed portfolio projects, two of them deployed live. Open to Data Analyst, BI Analyst and Operations Analyst roles in Germany from September 2026. ## Contact and profiles - Email: sergey@kasatov.de - Website: https://kasatov.de/ - LinkedIn: https://www.linkedin.com/in/sergey-kasatov - GitHub: https://github.com/sergey-kasatov - Tableau Public: https://public.tableau.com/app/profile/sergey.kasatov/vizzes This website is the primary node. The profiles above are secondary representations of the same person. ## Positioning Seventeen years inside vehicle development taught the habit that matters most in analytics: find the root cause, quantify it, and make the result something a non-technical stakeholder can act on. The work was always data work, it just ran on Excel pivots, KPI packs and 8D reports instead of Python and SQL. The strongest fit is analytics in manufacturing, quality or operations, where the domain knowledge is not decoration: knowing why a brake-condition signal dominates a failure model, or what a launch-phase issue backlog actually costs, is the difference between a model that scores well and a model someone trusts. ## Skills Technical: SQL, Python, pandas, Tableau, Excel and Power Query, machine learning, NLP, deep learning, AI agent workflows (n8n), prompt engineering and structured LLM outputs, AI strategy and governance, data cleaning, exploratory data analysis, KPI design, dashboards, DBeaver, Jupyter. Domain: automotive quality engineering, 8D problem solving, SPC, PPAP, APQP, launch management, body in white and closures, supplier quality. Languages: Russian (native), English (C1), German (B2). ## Portfolio projects Eleven completed projects. Every metric below is reproducible from the linked repository. ### 1. FreshMart Demand Planner (live) Retail sales forecasting end to end: EDA, feature engineering, comparison of 10+ models across statistical, machine-learning and deep-learning families, HyperOpt tuning and MLflow experiment tracking. Champion: HyperOpt-tuned Random Forest, RMSE 138.0, MAE 93.1, MAPE 21.3 percent, R2 0.43 on a chronological Jan-Mar 2014 test window. Deployed as a Streamlit planning tool that speaks planning language rather than metrics. - Live app: https://freshmart-demand-planner.streamlit.app - Code: https://github.com/sergey-kasatov/time-series-forecasting-project - Case study: https://kasatov.de/case-freshmart.html ### 2. Predictive Maintenance: fleet risk and cost Three business questions on a 92,000-row logistics fleet dataset: which vehicles need maintenance (classification, PR-AUC 0.902 with XGBoost, F1 0.784 with logistic regression), what it will cost (regression, R2 0.885 and RMSE 530 USD with Random Forest), and how to segment the fleet by risk (KMeans k=2, silhouette 0.083, interpreted post hoc as overloaded versus normally loaded vehicles). SHAP identified Brake_Condition as the dominant risk predictor, confirming domain expectation. - Code: https://github.com/sergey-kasatov/predictive-maintenance-ml - Case study: https://kasatov.de/case-maintenance.html ### 3. NLP: disaster tweet classification Shipped an interpretable TF-IDF plus logistic-regression classifier at accuracy 0.819, benchmarked against a CNN trained from scratch and a fine-tuned DistilBERT (0.839). The explainable model won the trade-off: every flag can be justified word by word, which matters more than the last two points of accuracy. - Code: https://github.com/sergey-kasatov/nlp-disaster-tweets-classification ### 4. CIFAR-10 transfer-learning study Controlled ablation with ResNet50 isolating the effect of data volume, augmentation, learning-rate scheduling and partial unfreezing, individually and combined. Accuracy improved from 66 percent to 82.7 percent across 60,000 images. Grad-CAM shows what the model actually attends to; the same mechanism drives automated visual inspection in manufacturing. - Code: https://github.com/sergey-kasatov/cifar10-resnet50-transfer-learning-experiments ### 5. TravelTide: customer segmentation Retention targeting for a travel platform. A SQL pipeline in Databricks aggregated 49,211 sessions into 30+ behavioural features for 5,998 users; a rule-based perk-fit score assigned each user the most relevant reward, producing five balanced segments. KMeans and DBSCAN served as benchmarks; the interpretable approach won on stability and deployment readiness. - Code: https://github.com/sergey-kasatov/traveltide-segmentation-retention - Presentation: https://youtu.be/jfZfYKk5VJ4 ### 6. Retail KPI Dashboard (live) Interactive Tableau dashboard for revenue, profit, discount and orders with drill-down by region, product category and period across roughly 10,000 rows, 48 months and 49 states. LOD expressions and dashboard actions give non-technical users self-service analysis, the same logic as a quality trend board on a plant floor. - Live: https://public.tableau.com/views/RetailKPIDashboardSalesProfitDiscount/MainDashboard ### 7. IATF Quality Report Prompt A prompt that drafts German ISO 9001 / IATF 16949 quality reports and is forbidden to derive any figure it was not given: a QMS report is a controlled record, and a derived figure is an audit finding even when the arithmetic is right. A negative-control test withheld one figure while leaving both operands in place as bait. Across 7 runs the prompt never computed the withheld figure and never introduced any figure absent from the input; the gap marker appeared in both required places in only 5 of 7 runs, and that weaker number is published as found, with the prompt deliberately left unfixed. A stdlib-only Python checker judges each run mechanically, and the prompt accepted a customer 8D form with no edit. - Code: https://github.com/sergey-kasatov/iatf-quality-report-prompt ### 8. Car Price Estimator Used-car pricing on 56,244 listings that returns a calibrated price range rather than a single number. Two of the three required model families (linear regression, decision tree) lost to a dealer-style lookup table; XGBoost beat it at MAE 1,016 USD and 11.9 percent median error against the lookup's 13.3, with the real gain on cars the lookup had never seen: 46 percent better there, 12 percent on familiar ones. Mileage is 2.2 percent of permutation importance against 62.5 for age, so mileage barely predicts price once age is known. Prediction intervals use split conformal calibrated per price band, after the global version hit its 80 percent coverage target on average while covering only 51.5 percent of cheap cars. Ships a CLI and an importable module that reproduces the notebooks exactly. - Code: https://github.com/sergey-kasatov/car-price-estimator ### 9. AI Order Triage Agent A 26-node n8n workflow that triages free-text wholesale orders in English and Portuguese. A schema-bound extraction step turns prose into structured items, deterministic Airtable lookups gather every fact, and one bounded LLM agent makes the single judgment a rule cannot express, choosing between standard, needs_review and needs_clarification. Logic routes and logs every order, and a human stays the sender: the workflow never contacts a customer. Across three runs of each of the ten sample orders, all 30 recorded classifications matched the course answer key and no input ever produced two different classifications. 2 of 8 processing steps are AI. Built on a fictional wholesale case, Volta Coffee Roasters in Porto. - Code: https://github.com/sergey-kasatov/llm-order-triage-agent ### 10. FX Random Walk Benchmark A walk-forward benchmark that asks the question the banking brief did not: compared to what? Twenty years of daily SGD/USD and CNY/USD, the naive "tomorrow equals today" baseline established first, then ARIMA, SARIMA and gradient boosting scored against it over 42,924 forecasts from 1,022 origins at three horizons. No model beat the baseline at any significance level under Diebold-Mariano with the Harvey correction, AIC selected ARIMA(0,1,0) for SGD, which is the random walk itself, and a purpose-built move_ratio measure explains every model's error at r = 0.97. The no-leakage property is proved mechanically rather than asserted: rewriting every observation after the cut date must leave the earlier features byte-identical. Results are priced against a USD 10m exposure. The only negative-result study in the portfolio. - Code: https://github.com/sergey-kasatov/fx-random-walk-benchmark ### 11. Neuron Platformer RL A 2D platformer built from scratch as a custom Gymnasium environment, with a level generator whose every gap is capped by the exact jump-physics envelope: an audit re-proves solvability across 1,200 seeds, finding 0 impossible transitions in 11,619. A PPO agent reading raw 84x84 pixels, with no coordinates, velocities or object lists, beats the hand-crafted 19-feature state-vector baseline 80 percent to 27 percent on 200 held-out seeds; a difficulty curriculum then carries it to 52 percent on medium and 55 percent on hard. Grad-CAM over the policy CNN shows attention on platform edges, gaps and enemies. Four classic RL failure modes are documented rather than hidden. A fourth phase replaces the frame stack with an LSTM reading one frame at a time: recurrence does learn the game but lands behind the frame stack on every tier, 29.5 percent on medium against the curriculum agent's 52. Getting the numbers honest took three extra experiments: a budget doubling that refuted the write-up's own excuse (typical performance rose, the peak did not), a noise analysis showing a 30-episode evaluation carries about plus or minus 14 points, and a 200-seed re-measurement of all four models that replaced every headline figure - the state baseline's published 50 percent was really 27. The write-up was rewritten rather than left flattering. - Code: https://github.com/sergey-kasatov/neuron-platformer-rl ## Experience ### SEGULA Technologies - Automotive Consultant Apr 2014 - Feb 2025, Cologne, Germany. Long-term assignments at Stellantis, Opel and Ford. **Project Engineer, BIW and Door Components (Stellantis, Ruesselsheim)**, Jul 2022 - Feb 2025. Led development-to-launch integration for about 100 body-in-white parts and coordinated about 150 engineering changes through ramp-up. A weekly KPI and quality trend pack tracking 10 to 15 top issues per week cut issue time-to-close from about 4 weeks to about 3 weeks and launch-phase rework by about 25 percent, delivering a 100k EUR cost saving. **Launch Quality Specialist, TSQ (Opel Plant Eisenach)**, Oct 2020 - Jun 2022. Launch quality for Opel Grandland X Facelift serial production: tracked vehicles through assembly sequences, analysed and resolved line-side issues, documented corrective and preventive measures. **Manufacturing Engineer, Launch and Logistics (Opel Eisenach and Ruesselsheim, Stellantis Trnava)**, Jul 2019 - Sep 2020. Material flow analysis for the Opel Grandland X PHEV introduction, Supermarket and E-Kanban optimisation, assembly concepts for the Opel Astra line integration, support for PSA process introduction on the Peugeot 208 and e208 BEV. **Group Leader, Body Exterior and Interior (Ford)**, Apr 2014 - Jun 2019. Supported 7 vehicle launches with a structured open-issues cadence across 3 plants and 5 suppliers. Led a team of 3 engineers. Created a weekly KPI pack for 6 stakeholders tracking SOP readiness on 70+ parts. Delivered 250k EUR in cost savings through standardisation and localisation initiatives. ### General Motors - Product Engineer Jan 2013 - Apr 2014, Saint Petersburg, Russia. Localised 25 BIW stamped parts through local steel and metal-forming suppliers, delivering 75k EUR in savings. Built a localisation tracker and weekly status reporting across Chevrolet, Opel and Cadillac programs. Owned SOP readiness for 2 vehicle programs. ### Earlier - Ford Motor Company, internship, Oct 2011 - Sep 2012, Cologne. Prototype-build readiness across supplier scope for the Ford Kuga (45 components); UHSS body structure repair methods. - Continental AG, internship, Mar 2011 - Sep 2011, Ingolstadt. Thermo-mechanical simulations for injection-moulded pressure distributors, cutting cooling time by 35 percent. - OJSC SOLLERS, Senior Supplier Quality Engineer, Jul 2008 - Sep 2010, Yelabuga. Audited 20+ suppliers, led PPAP and APQP for 35 parts, supported launches for 3 vehicle models on FIAT and ISUZU programs. - AO PO ElaS, Process Engineer, Feb 2008 - Jul 2008, Yelabuga. Developed and validated 8 CNC programs; reduced manufacturing costs by about 15 percent. - OJSC KMPO, intern, Aug 2007 - Dec 2007, Kazan. ## Education **Master School Institute of Technology** - Data Analytics Program (AZAV), Advanced Data Science and AI. Sep 2025 - Sep 2026, full time. Python, SQL and pandas on large datasets; Tableau and Matplotlib for BI reporting; statistical analysis and predictive modelling; time series, computer vision and NLP with LLMs; AI-assisted analysis workflows, AI strategy and governance, and automation with agent concepts. **Technische Hochschule Ingolstadt** - International Automotive Engineering, Oct 2010 - Mar 2013. Master's program, coursework completed, degree not awarded. ## Certificates Nine certificates with verification links are listed on https://kasatov.de/about.html. The most recent is "AI Agents PRO" (Jul 2026), which covers agent workflows; the personal knowledge base that AI agents work in daily is the practical result. ## Availability Open to Data Analyst, BI Analyst and Operations Analyst roles in Germany from September 2026. Based in Cologne. On-site, hybrid or remote. ## About this site Static HTML, no cookies, self-hosted fonts. The only external request is the cookieless GoatCounter visitor counter (no personal data stored, no consent banner required). Available in English and German (https://kasatov.de/de/index.html). Written and maintained by Sergey Kasatov.