BC
Résumé Email
Portrait of Bohdan Chuprynka

Bohdan Chuprynka

AI Engineer

I build AI products that survive contact with reality.

Email Résumé GitHub LinkedIn

Experience

Founding AI Engineer · Remly Retrieval, agents, and product features across a Next.js web app and a Swift macOS client: local files and connected apps searchable by meaning, with cited answers. Jun 2026 – Present
  • Built exact-filename search, relevance ranking, and Ask context with citations across web and Mac.
  • Shipped connector integrations, Mac ingestion batching, and OAuth 2.1 MCP retrieval with access controls and cross-platform tests.
TypeScriptNext.jsSwiftPostgreSQLpgvectorMCPOpenAI API
Cybersecurity Intern, Data Protection · Cleveland Clinic Reviewed 17 sensitive-data classifiers, engineered 13 custom classifiers, and built a privacy-safe validation harness with six-format synthetic tests. Jun – Aug 2026
  • Regex classifiers for privacy and compliance workflows, tuned on synthetic data with holdout sets.
  • Synthetic evaluations reached approximately 100% precision and recall across six file formats.
PythonRegexData protectionSynthetic evaluation
Co-Founder & CTO · Aximon Owned backend and AI infrastructure for an AI coding mentor. 1.7K extension downloads, 1,000+ beta testers, 6 partnerships. Mar – May 2026
  • Built tutoring and platform features across a TypeScript extension and backend: learner and code-context injection, Python validation, automated checks, interactive IDE previews.
  • Model-provider routing, shared LLM-output caching, cost controls, cloud-synced chat, and editor analytics.
TypeScriptNext.jsNestJSPostgreSQLpgvectorBullMQ
AI Engineer Intern · Cleveland Clinic Built an nnU-Net vertebral segmentation pipeline for spinal X-rays: preprocessing, training, post-processing, and evaluation with clinical collaborators. Jun – Aug 2025
  • Customized nnU-Net planners, loss functions, and post-processing filters for vertebral structures.
  • Validated outputs with radiologists and neurosurgeons for surgical planning, without exposing protected data.
PyTorchnnU-NetMedical imagingPython

Projects (8)

Remly.it homepage

Remly

Product · current

The context layer for Mac and AI. Search and Ask across local files and connected apps, with citations. I build the retrieval, cited Q&A, connector, and cross-platform layers.

Role
Founding AI Engineer
Surface
Next.js web + Swift macOS
remly.it
Persona-RAG side-by-side comparison of the local model and the API

Persona-RAG

Research + product

Can a free local fine-tune match a paid RAG stack on voice? A Qwen2.5-3B LoRA against a gpt-4o-mini hybrid-retrieval product, scored on a leak-free hold-out after I caught a 90% train/test leak in my own eval.

Cost
$0 local vs $0.37 per 1k replies
Hold-out
n = 300, ~11× fewer tokens
GitHub Paper
Gallery · 4
Live comparison: the cloud API on the left, the local model on the right
Live comparison, real model output
Headline result chart
Headline result, 95% intervals
Voice-register metrics chart
Register tics vs my real baseline
Retrieval leak guard chart
Leak guard, 28% removed to 0
Aximon adaptive Python tutor interface

Aximon

AI product

The AI foundation for an adaptive Python tutor: multi-step tutoring, pgvector retrieval, model routing, evaluation, and asynchronous course generation.

Downloads
1.7K
Beta testers
1,000+
Partnerships
6
Gallery · 3
Theory lesson with the tutor
Theory lesson + tutor
Practice feedback screen
Practice feedback
Curriculum sequencing tree
Curriculum sequencing
Policy RAG system architecture

Policy RAG

Retrieval

Citation-grounded Q&A over policy documents. The original Python/NumPy app uses hybrid dense and BM25 retrieval without LangChain behind FastAPI; the live demo is a Sites port of that flow.

Retrieval
Dense + BM25, from scratch
Shape
FastAPI · Redis · Docker · CI
Live DemoGitHub
Vertebral segmentation overview Vertebral segmentationnnU-Net pipeline for spinal X-rays, built at Cleveland Clinic. Private work, shown without protected data. Medical imaging · 2025
Sales forecasting results Sales forecastingFive model families across 3M+ retail rows and 54 stores. Best R² 0.83. GitHub Sentiment analysis model comparison Sentiment analysisClassical ML against deep learning on 1.6M tweets. Best accuracy 80.6%. GitHub Tableau business dashboard Business dashboardInteractive sales, margin, and regional performance in Tableau. Tableau

Skills

01LLM apps + retrievalRAG, hybrid search, agents, model routing, fine-tuning LangGraphHugging FaceQdrantpgvectorBM25LoRAOpenAI APIMCP
02Production AI + evaluationBackend services, test harnesses, observability, cost and latency control PythonFastAPITypeScriptNext.jsPostgreSQLDockerMLflowpytestGitHub Actions
03Machine learningDeep learning, computer vision, NLP, model training PyTorchTensorFlowscikit-learnnnU-NetOpenCV
04Data + analyticsPipelines, SQL, evaluation, dashboards PandasNumPySQLRedisTableau

Let's talk

Roles, offers, contracts, or just a thought. I'm open to all of it.

bohdan.chuprynka@gmail.com
GitHub LinkedIn Résumé