Danylo Sovgut
Danylo Sovgut
New York, NY/2026

Meet Danylo Sovgut.An astrophysicist turned AI engineer.

I build retrieval pipelines, evaluation systems, and inference infrastructure for ML products that ship to real users — not just notebooks.

Scroll

01 /

Selected Work

2025· In production · BMG

ML pricing pipeline for sync licensing

Two-stage classifier + soft-routed regression, invented after standard approaches (quantile regression, conformal prediction, Bayesian quantile) failed on the long-tailed distribution. Shipped as a FastMCP server with Shapley-value explanations, alongside four other tools the licensing team uses in their internal AI chat.

PythonXGBoostLightGBMFastMCPSHAPSpotify APIGeminiAim
View case study
Reduction in MAE vs the legacy rate card
~78%
Projected annual revenue uplift
$7M+
MCP tools in production for the licensing team
5
2025· In production · BHI

Integrated AI ops stack for a VC firm

Connected six previously disconnected tools (Attio, Notion, Granola, Box, Gmail, Slack) into one Claude-driven workspace. Eight Claude Projects + six Cowork automations + a $0/mo Python news monitor.

ClaudeMCPAttioNotionGranolaBoxGitHub Actions
View case study
Solo build, end-to-end
~1 month
Custom Claude Projects in production
8
Custom news monitor infra cost
$0
2025· In production · BHI

Investment memo writer for VC partners

Multi-source ingest (CRM + Notion + Tavily + uploaded decks) feeds a single-call Gemini Pro generator. Core rule: the writer never sees the raw corpus — only user-verified facts.

FastAPIReactGemini ProPydanticPostgresTavily
View case study
Faster first-pass memo prep
~24×
Memo sections, single-call generation
12
Used by BHI partners on every deal cycle
Daily
2025· In production · PRH

Book tagging for retail ad selection

GPT-5-mini grounded on SerpAPI search. K-means surfaced six demand archetypes; HDBSCAN sub-clustering caught a hidden 5.28× ROAS pocket K-means averaged out; SHAP showed which tags actually moved performance.

GPT-5-miniDatabricksSnowflakeSerpAPIscikit-learnHDBSCANSHAP
View case study
Books tagged in production
~13K
ROAS pocket (n = 12) via HDBSCAN
5.28×
Candidate ad titles surfaced
189

02 /

Open Source


03 /

Experience

── Education

Sep 2024 — Dec 2025

University of Chicago

MS, Applied Data Science

4.0 GPA

Sep 2017 — May 2020

University of Illinois at Urbana-Champaign

BS, Astronomy & Physics

3.5 GPA · Honorary List

── Roles

Aug 2025 — Present

Bertelsmann Future Leaders Program

Sole engineer · BMG · BHI · PRH
  • Built production ML and AI products across three Bertelsmann subsidiaries — sole engineer on each, owning stakeholder discovery, architecture, build, and roll-out.
  • BMG: ML pricing pipeline for sync licensing — ~78% MAE reduction, projected $7M+ annual uplift. Shipped as a FastMCP server with Shapley-value explanations, alongside four other MCP tools the licensing team uses in their internal AI chat.
  • BHI: integrated AI ops stack + investment memo writer for the partner team — full stack designed and shipped solo in roughly one month; used daily by partners on every deal cycle.
  • PRH: book-tagging pipeline for retail ad selection — gathered new data and successfully defended a year-old initiative to the Head of Data Science within a month.
Sep 2024 — Dec 2024

University of Chicago — Graduate Projects

Three end-to-end AI products
  • Global-IQ agentic AI assistant — replaced a legacy form calculator with an MCP-integrated frontend for personalised global-mobility recommendations.
  • UChicago Data Science RAG chatbot — scraped program content into transformer embeddings stored in a vector DB; GPT-4 served via a Chainlit UI.
  • Multimodal e-commerce chatbot — CLIP image embeddings + LLM identify products from photos and answer follow-up questions; published on GitHub as Chatbot-for-Amazon.
Dec 2018 — May 2020

University of Illinois at Urbana-Champaign

Computational Astrophysics Engineer
  • Modelled near-Earth supernova and gamma-ray-burst rates; improved baseline model accuracy by ~50%.
  • First author on the resulting paper — “Cosmic Hazard Forecast: Rates of Near-Earth Supernovae and Gamma-Ray Bursts” (Sovgut, Krishnan, Trauth, Fields), published in MNRAS.
  • Selected from 800+ undergraduates to present at the Illinois State Capitol (Capitol Scholars Research Award).
Jun 2019 — Aug 2019

Main Astronomical Observatory, NAS Ukraine

Data Analyst & Observatory Operator
  • Automated satellite-tracking data ingestion and cleaning, cutting observation prep by 30%.
  • Analysed 300+ satellite observations for an international scientific consortium contributing to satellite-orbit study.
  • Co-presented a poster at an international science conference in Stuttgart, Germany.

04 /

Languages & Skills

── Languages

English
Fluent
Spanish
Fluent
Russian
Fluent
Ukrainian
Fluent
Portuguese
Intermediate

── Production-grade tools

ML / DS
Pythonscikit-learnXGBoostLightGBMCatBoostPyTorchTensorFlowSHAPHDBSCANPydantic
LLM / Agentic
OpenAI APIGemini APIAnthropic ClaudeMCPFastMCPVercel AI SDKRAGQLoRAPEFTStructured outputs
Computer Vision
OpenCVResNet50YOLOv11-poseCLIPQdrant
Data infra
BigQueryDatabricksSnowflakePostgresSQLTavilySerpAPISpotify APIAim
Web / App
FastAPIReactNext.jsTailwindVite
Cloud / Ops
Google CloudRailwayGitHub ActionsDockerGit

05 /

Pluto Academy

Apr 2020 — Present· Founder & Director

A non-profit online physics school
I've run for five years.

I founded Pluto Academy to make university-level math, physics, and astronomy accessible to students who otherwise wouldn't see it. Five years in, it's a small but real institution: a recorded curriculum and a long list of students who've gone on to STEM degrees.

It's the longest-running thing I've built — and the clearest signal of how I work when no one is asking me to.

Students taught
1,000+
Original lessons
70+
Hours of curated video
8+
5-star satisfaction
99%

06 /

Contact

Get in touch.

Email is the fastest channel.