Сергій Ветеран

Serhii Pedchenko
Python Engineer — Backend · AI / ML · Frontend
Kyiv, Ukraine · [відкрити контакти](див. вище в блоці «контактна інформація») · [відкрити контакти](див. вище в блоці «контактна інформація») · Telegram: @psg_sc · [відкрити контакти](див. вище в блоці «контактна інформація») ·
github.com/scorpion920

SUMMARY

Python engineer working across the whole stack of a data product: the ingestion layer, the models that run
on top of it, the API, and the interface the operator works in. My focus is production machine learning in
places where a wrong answer costs money, so I design for verification — measured quality, shadow rollout,
and a person in the loop where it matters. I have built and shipped two production platforms end to end as
the only engineer on both. Before engineering I spent eight years in commerce and procurement.
TECHNICAL SKILLS

Languages: Python 3.11, SQL, TypeScript, HTML, CSS

Frameworks: FastAPI, SQLAlchemy (async), aiogram

AI / ML: LangGraph, MCP, OpenAI API, local LLM inference (Ollama), RAG, embeddings, LightGBM, scikit-learn, CRF, pandas,
NumPy, SHAP
Databases: PostgreSQL, pgvector, Alembic, MinIO / S3-compatible storage

Caching & Messaging: Redis, Celery

Frontend: React 18, TypeScript 5, Redux Toolkit, TanStack Query, Vite, Chart.js

Cloud & DevOps: Hetzner Cloud (provisioning, networking, operations), GCP (Gmail API, OAuth), Docker, Docker Compose,
Linux, Nginx, HAProxy, Prometheus + Grafana, CI/CD, GitHub Actions
Other: pytest, Playwright, Telethon, pymorphy3

PROFESSIONAL EXPERIENCE

Full-stack Python & ML Engineer Oct 2025 — Present
Sole engineer on two production platforms — architecture, data, ML, backend, frontend, deployment and releases (client under
NDA)

Designed and built a multi-tenant analytics and forecasting platform. The team could cover about a third
of the portfolio by hand and estimated the rest; the platform covers effectively all of it, every day, one
entity at a time, and the cost of forecast error fell by roughly two thirds. The ingestion layer normalises
incompatible third-party data feeds into a single store with versioned revisions and idempotent
reprocessing, so a source that resends corrected history creates a new revision only where a value actually
changed. Several forecasting models compete per segment and are scored continuously against actuals,
so a model that drifts is taken out of service without anyone watching it. The operator UI covers
dashboards, time-series charts, alerting and anomaly detection, with human-in-the-loop confirmation on
anything that changes a plan.

Built a document-understanding pipeline that extracts structured terms from free-text contracts and
drives the downstream action end to end, without an operator: a case that took a specialist 1–3 minutes
now completes in seconds. Running several discovery channels in parallel instead of polling one cut time-
to-detection from a median of 143 seconds to about 3 — the bottleneck was never our code, it was
waiting for someone else to index the page. An ensemble of narrow NLP models each extracts one field
and returns calibrated probabilities with alternatives; an integrator searches those for an internally
consistent reading and abstains when the evidence is weak. A gradient-boosted model over 31
engineered features scored 91.8% against 64.4% for a fine-tuned 184M-parameter transformer on the
same task — seconds of training against four hours, on CPU — so the transformer was retired. Against
the reference set the rewritten pipeline matched 38 of 38 expected results with no wrong values; the
implementation it replaced matched 3 of 29. A new model ships in shadow beside the rule it would
replace and is promoted only after an automated evaluation gate. Trained against a corpus of 25,949
document pairs.
Both platforms run on cloud servers I provision and operate. Releases are tagged and pulled by each
installation on a schedule; migrations are expand-contract and a test fails the build on a breaking change,
so upgrades happen without downtime. I wrote the ownership model, the review process and the
onboarding document that let a second engineer join the work.
Stack: Python, FastAPI, SQLAlchemy (async), Celery, PostgreSQL, Redis, LightGBM, scikit-learn, CRF, LangGraph, MCP, React,
TypeScript, Docker, Prometheus / Grafana

Procurement automation — own initiative Apr 2025 — Oct 2025
Alongside the role below
▸ Built the first working version of the procurement monitoring system while still running the business
line. It went into daily use and became the platform above.

Head of Business Line Sep 2024 — Sep 2025
Company in the public-procurement sector, Ukraine
▸ Ran the department and a dedicated business line: process design, analytics, work with procurement
platforms and counterparties. This is where I learned the domain I later automated.

Military Service Aug 2022 — Apr 2024
Armed Forces of Ukraine

Sales Manager → Head of Sales → Head of Procurement 2014 — 2022
Private company, food-supply sector
▸ Eight years at one company, from manager to head of procurement. Teams of 5 to 22 people: hiring,
training, workload allocation, performance metrics, volume and budget planning.
EDUCATION

Master's Degree, International Economic Management — National University of Life and Environmental
Sciences of Ukraine, 2014
Python Developer — CyberBionic Systematics and N-Code, certified, 2024 — 2025
LANGUAGES

Ukrainian — Native. English — Upper-Intermediate (B2).