Особисті дані приховані

NAZAR LATKO
Data & AI/ML Engineer
[відкрити контакти](див. вище в блоці «контактна інформація») • [відкрити контакти](див. вище в блоці «контактна інформація») • Kyiv, Ukraine •
PROFESSIONAL SUMMARY
Data & AI/ML Engineer delivering production AI systems, backend services, and data pipelines with Python,
SQL, APIs, async workflows, and cloud infrastructure. Applied ML background in computer vision, anomaly
detection, and tabular modeling; experienced from R&D and architecture to deployment, with a focus on
reliability, data quality, and business impact. Selected among the faculty's top 20 students at ML Week 2025
and 2026.

WORK EXPERIENCE
Atschool | AI/ML Engineer | Nov 2025 – Present [Full-Time]
• LessonLM Video - AI-Powered Presentation-to-Video Platform:
◦ Sole developer of a PPTX-to-MP4 platform, owning R&D, architecture, full-stack delivery, and production
integration; cut video creation time from 1–2 hours to ~10 minutes.
◦ Built a custom PPTX parser and structured animation pipeline, combining LLM-planned effects, Whisper
timing, and Remotion rendering to synchronize scenes with narration.
◦ Built a Next.js/React/TypeScript frontend and FastAPI/Celery/Redis backend with generation
checkpoints, failure handling, and back-office integration.
◦ Added in-product feedback, screenshots, and logs for debugging and improvements after launch.
• LessonLM — AI-Powered Educational Content Platform:
◦ Built an LMS-integrated AI content platform generating assessments, notes, glossaries, and visuals
automatically attached to lessons.
◦ Engineered async document processing and generation with FastAPI, RabbitMQ, Celery, PostgreSQL,
S3, retries, and AI-provider failure handling.
◦ Added RAG (text extraction, chunking, embeddings, retrieval) for reusable 400+ page textbooks,
reducing estimated LLM processing costs by ~80–90%.
• Course Recommendation Engine — Hybrid AI-Assisted Recommendation System:
◦ Built an explainable course recommender for academic improvement, advanced study, NMT prep, and
interests, ranking by grades, skills, preferences, and enrollment history to support upselling.
◦ Addressed sparse-data cold start with deterministic ranking and periodic Gemini course/skill
classification; cached precomputed recommendations to avoid runtime LLM calls.
• Student Performance Prediction & Explainable Learning Analytics:
◦ Trained LightGBM to predict students' average grades; used XAI to explain feature contributions and
factors lowering predicted results.
◦ Built a lesson difficulty ranking system using aggregated student grades, time spent, activity and
question-level performance to identify issues
• Automated Essay Scoring — LLM Evaluation & Human-Aligned Assessment:
◦ Built an LLM essay scorer (0–100) to automate preliminary assessment and reduce manual grading.
◦ Evaluated AI scores against teacher-graded reference essays with Langfuse and MAE; refined prompts
to improve alignment with human assessments.
• Telegram Chat Analyzer — LLM-Powered Communication Quality Monitoring:
◦ Developed an LLM-assisted quality monitoring pipeline for mentor–student chats across Telegram and
the learning platform, detecting delayed responses, recurring complaints, sentiment, and communication
issues.
◦ Integrated Gemini-based text analysis and Langfuse observability to support systematic review of
communication quality.
• UK EdTech Tender Intelligence — Multi-Source Data & AI Pipeline:
◦ Built a multi-source UK procurement/education data pipeline with API ingestion, normalization,
deduplication, caching, and Gemini-powered semantic relevance filtering.
◦ Delivered a searchable, weekly refreshed dataset of 375 relevant tenders, 15,223 schools, 3,874 trusts,
51,196 contacts, and 7,220 email leads; added optional Jev classification to improve efficiency.
• Data Analysis:
◦ Audited two years of student/payment data, identified traceability gaps, and proposed structured tracking
for more reliable financial analytics.
◦ Analyzed registration-to-payment conversion, identified onboarding friction, and proposed improvements
to value communication and payment flow.
◦ Analyzed user behavior, payments, and operational data, uncovering anomalies and bottlenecks to
inform product improvements.
Logika School | Programming Tutor | Aug 2024 – Feb 2026 [Part-Time]
• Taught programming fundamentals to children and teenagers, simplifying complex concepts.
• Adapted teaching to different learning levels, strengthening mentoring and communication skills.

EDUCATION
Bachelor's Degree in Applied Mathematics | GPA: 93/100
National Technical University of Ukraine 'Kyiv Polytechnic Institute' | Sep 2023 – Jul 2027
Certificates: link
PROJECTS & COMPETITIONS
Unified Web Parser — Schema-Driven AI Web Data Pipeline
• Built a schema-driven Python web-to-dataset engine with Crawl4AI/Playwright and GPT Decisions for
constrained navigation, extraction routing, and JSON-LD/DOM/LLM parsing.
• Implemented validation, provenance, deduplication, incremental crawling, checkpoints, and billing-aware
recovery; supported SQLite/PostgreSQL and Parquet/JSONL/CSV exports.
• Reduced estimated AI token costs by 93.7% on a five-page benchmark (52 quality checks preserved); built a
FastAPI/SSE dashboard for live runs, records, and usage.
ATLAS — Self-Hosted Agentic AI System | Hermes, MCP, Oracle Cloud
• Deployed a persistent Hermes agent on Oracle Cloud Linux with Telegram, memory, scheduled jobs,
browser, terminal, and code-execution tools.
• Extended Hermes with modular skills, MCP integrations, multi-step research/development workflows, and
persistent gateway services for unattended operation.
• Built an MCP-driven news-to-video workflow from research and storyboards to asset selection and
programmatic Remotion rendering.
Grimoire - Interactive AI Notebook
• Architecting a stylus-first AI notebook with handwriting event streams, replayable sessions, work
provenance, and AI analysis of problem-solving steps.
• Designed a subject-agnostic Rooms → Notebooks → Works → Pages model with adapters for
cross-disciplinary AI review.
Industrial Anomaly Detection & Edge Optimization - ML Week 2026
• Benchmarked Student–Teacher, EfficientAD (Anomalib), and PatchCore on MVTec AD; selected EfficientAD
for detection/latency trade-offs, reaching F1 0.99 (Grid), 0.98 (Bottle), and 0.94 (Carpet/Metal Nut).
• Optimized edge inference with INT8 quantization: 4–7× faster CPU (160–217 → 30–40 ms); benchmarked
FP32 GPU at 55 FPS (18.1 ms/frame).
HACKATHONS & AWARDS
• INT20H 2026 - 1st Place: Anomaly detection analysis and acceptance rate forecasting.
• ML Week 2026 (GlobalLogik Track): 1st Place in category of Industrial anomaly detection.
• Diia.AIContest - Participant: AI Agent for automated completion of government documents.
• Skellar X Monobank - Participant
• INT20H 2025 - 6th Place: Fraud detection and time series forecasting.
• Brainstack Data Camp - Participant
TECHNICAL SKILLS
Languages: Python, JavaScript / TypeScript, SQL
ML / AI: PyTorch, Scikit-learn, GBDT, CNN, RNN, Transformers, XAI, CV, NLP, anomaly detection
MLOps / Evaluation: MLflow, Langfuse, ML Inference
LLM & Agents: OpenAI, Claude, Gemini, Qwen, OpenRouter, RAG, MCP, agentic workflows,
prompt/context engineering, Ollama.
Backend & Infra: FastAPI, Flask, Celery, RabbitMQ, BullMQ, Redis, Docker, CI/CD
Data & Storage: PostgreSQL, MongoDB, Redis, ChromaDB, Qdrant, S3, Pandas, NumPy, BigQuery
Frontend / Media: React, Next.js, Remotion, Hyperframes, Lottie, TypeScript
Cloud & Tools: Azure (basic deployment), Git, Postman, Jupyter Notebook
Spoken Languages: Ukrainian (Native), English (B2)
AI Dev Tools: Claude Code, Codex, Devin, Cursor, GitHub Copilot, Perplexity