📌 Чому я підходжу (дзеркало JD)
- "Prompt Engineering, RAG, LLM pipelines — production" — маю в production: ZRT classifier (structured prompt + JSON schema output), autoblog pipeline (Anthropic/OpenAI), Vision QA (multi-lens screenshot analysis), IMPRINT (self-learning failure classifier).
- "Eval datasets, A/B тести промптів, regression-тести" — PILOT framework: JSONL failure база + per-site regression testing + IMPRINT self-learning (саме це і є "eval pipeline").
- $2500–4000, власні продукти Galaktica" — мені цікаво будувати AI всередині продукту, не консалтинг.
- "Python, FastAPI" — Python є (scraper pipeline, Vision AI scripts, Telethon), FastAPI не primary але розумію REST API patterns (NestJS).
🎯 Summary
Я інженерю LLM pipeline, prompt systems і RAG архітектури в production. Зараз маю 3 активних AI pipeline: ZRT Classifier (LLM з structured JSON output класифікує пошукові запити), Vision QA Pipeline (скриншот → multi-lens LLM аудит), IMPRINT (self-learning failure classifier — накопичує JSONL failures, per-site patterns, 0 false positives). Також: autoblog (OpenAI/Anthropic для SEO контенту) і scraper Vision AI (brand detection для ~10K+ listings). Python workflow — щоденно для scraper pipeline і Vision AI scripts. Backend: NestJS + TypeScript (primary). FastAPI — знаю REST patterns, можу освоїти швидко.
🚀 LLM / AI Engineering Portfolio
IMPRINT — Self-Learning Failure Classifier
Production: PILOT framework
# Conceptual flow (TypeScript implementation)
classify_failure(screenshot, html, console_log) → {
category: "auth_error" | "layout_shift" | "timeout" | ...,
confidence: 0.95,
action: "retry" | "skip" | "delete" | "retry_proxy"
}
- JSONL failure база накопичується per-site
- LLM classify once → cache → rule-based fallback (0 Vision API calls after warm)
- A/B тест: LLM-only vs IMPRINT cached → IMPRINT: 25→0 Vision API calls/run
ZRT Classifier — Structured LLM Output
Production: AutoShara + PGT
// Prompt engineering з JSON schema enforcement
classifier.classify(searchQuery) → {
listingType: "SALE" | "RENT",
subIntent: "buy_domestic" | "rent_longterm" | ...,
confidence: number,
metadata: { brand?, city?, priceRange? }
}
- Structured output: JSON Schema validation, fallback on malformed response
- Per-vertical prompts: SALE prompts ≠ RENT prompts (different training context)
- Eval dataset: A/B між Claude і GPT-4 classifier prompts
Vision AI Pipeline — Photo Quality + Brand Detection
Production: AutoShara scraper (~10K+ listings)
# Python pipeline
for listing in batch:
screenshot = capture_listing_photo(listing.url)
result = openai_vision.analyze(screenshot, prompt=BRAND_DETECTION_PROMPT)
if result.brand != listing.brand:
flag_for_correction(listing, result)
- Batch processing з rate limiting
- WARP proxy для production (server IP blocked on AutoRia)
- Результат: ~80% data quality improvement
Autoblog Pipeline — LLM Content Generation
Production: AutoShara + PGT
- Anthropic Claude SDK: structured prompts → SEO article generation
- Variables: location, vertical, keyword targets → personalized content
- Pipeline: trigger → generate → validate (length, keywords) → publish → GSC submit
🛠️ AI Engineering Stack
| Категорія | Технології |
|---|---|
| LLM APIs | OpenAI (GPT-4, Vision, embeddings), Anthropic Claude (claude-3-5-sonnet) |
| Prompt Engineering | Structured JSON output, few-shot, chain-of-thought, multi-lens |
| RAG | Knowledge Pack per role, IMPRINT JSONL база, context injection |
| Python | Playwright, Selenium, Vision API scripts, Telethon, requests |
| TypeScript/Node.js | NestJS (primary backend), LLM integration в production |
| Eval | JSONL failure datasets, A/B prompt comparison, regression testing |
| AI Dev Tools | Claude Code (primary), GitHub Copilot, Cursor |
| Infrastructure | PostgreSQL, Redis, PM2, Hetzner VPS |
💡 LLM Engineering Patterns
Prompt Engineering
- Structured output: JSON schema enforcement, validation layer, malformed fallback
- Few-shot examples в classifier prompts для higher accuracy
- Multi-lens pattern: один скриншот → 3 різні prompt-lens (UX/Design/Architecture)
Eval & Regression
- JSONL failure база: накопичується автоматично, version-controlled
- A/B: порівнюю Claude vs GPT-4 на однаковому eval dataset
- IMPRINT: classify once → cache → правило → 0 API calls на повторних патернах
Cost Optimization
- IMPRINT: 25→0 Vision API calls/run → ~$0.50/month замість $12/month
- Batch processing з rate limiting
- Cache frequently-seen patterns
📌 Чим відрізняюся
- Production LLM pipelines — не toy notebooks
- IMPRINT self-learning classifier = досвід eval + optimization
- Python (scraper, Vision AI) + TypeScript (NestJS) — hybrid stack
🎓 Background
Self-taught AI engineer. 2+ роки production LLM pipeline delivery.
🌍 Мови
Українська: рідна | Англійська: B2+ | Російська: пасивна
📎 Посилання
Продукти: https://autoshara.com | https://pgthub.com LinkedIn: linkedin.com/in/konstantyn-onosov-9a8929403
