Abonyi János
Software Developer & AI Engineer

A LangGraph-orchestrated Corrective-RAG (CRAG) pipeline that audits knowledge-base and bot responses for adherence to legal, GDPR/HIPAA privacy, and company policy. It screens input for prompt-injection and jailbreaks, redacts PII before anything reaches the LLM, grades retrieved documents, and falls back to web search when local context is insufficient.

• Fail-safe ordering: input guardrails → PII redaction happen before text reaches embeddings or the LLM.
• Two-stage input screening — cheap regex first, LLM classifier second.
• CRAG correction: documents are graded; weak local context triggers a Tavily web search before generation.
• Grounded generation as a structured ComplianceAudit, re-checked for grounding/toxicity with one bounded self-correction.

An AI-powered HR platform that unifies five database paradigms — Relational, Graph, Document, Vector, and Big Data — behind one orchestration layer. An LLM uses function calling to route natural-language questions across PostgreSQL, Neo4j, Firestore, Pinecone, and BigQuery ML, chaining results to answer questions no single store can.

• Multi-hop GraphRAG: e.g. 'find a great listener mentored by someone who knows Python' chains Pinecone → Neo4j → PostgreSQL → BigQuery ML.
• Transactional outbox pattern keeps five stores consistent without a distributed commit — business row + event commit atomically in one PG transaction.
• At-least-once delivery with idempotent handlers (MERGE / upsert), dead-letter queue, and a drift detector across all stores.
• BigQuery ML logistic-regression model predicts employee flight risk directly in SQL.

Upload PDF / DOCX / TXT study materials — including scanned PDFs — and get instant, source-cited answers to any question, including A/B/C/D multiple choice. Each subject lives in its own isolated knowledge base with full Q&A history.

• Retrieve top-20 from Pinecone, rerank to top-5 with Cohere, then answer with GPT-4o at low temperature for grounded responses.
• 3-stage PDF extraction with OCR fallback (pdfjs-dist → pdf-parse → Mistral OCR) for fully scanned documents.
• Per-knowledge-base Pinecone namespaces for full document isolation.
• Answers rendered as Markdown + LaTeX (KaTeX), always in the question's language.

A cloud-native photo app: every uploaded image is auto-analyzed by Google Cloud Vision, which returns content labels, so you can search 'dog', 'mountain', or 'receipt' with no manual tagging. Built to compose three Google Cloud services behind one Flask app.

• Upload → Cloud Storage, analyze → Vision API labels, store metadata → Firestore, all in one pipeline.
• Case-insensitive label search via a lowercased labelsLower array and Firestore array_contains.
• Containerized with Docker and Gunicorn, deployed to Google Cloud Run with identity-based auth.

An ML project that analyzes Formula 1 telemetry (speed, throttle, brake, RPM, gear) to distinguish driving styles between drivers and build a predictive model that identifies a driver from a lap. Data is pulled per-lap with FastF1, distance-resampled into uniform time series, and turned into per-lap feature matrices.

• Telemetry ingestion via FastF1 with local caching across Grand Prix sessions.
• Distance-based resampling to align laps into comparable, uniform feature vectors.
• Exploratory analysis (box/violin/Q-Q plots) comparing throttle, brake and speed distributions per driver.
• Supervised classification model to attribute laps to drivers (VER, RUS, NOR, PIA).