Generative AI Creative& Brand Filmmaker

A single-station industrial QC agent that pairs a PatchCore vision detector (0.999 image AUROC on MVTec AD, CPU-only) with a LangGraph plan → act → observe loop. It reads machine, batch, and operator history from a SQLite MES to judge whether a fault is random or systematic, decides pass/rework/reject as a deterministic, auditable function, and escalates low-confidence cases to a human. A drift monitor (OOD gate + PSI) guards the pass path against camera and lighting shift. The LLM only writes the narrative; 88 tests cover the rest.

A LangGraph pipeline of four agents — Monitor → Diagnostics → Recommendation → Workflow — that watches industrial pumps, flags anomalies with ISO 10816 vibration zones and z-score analysis, and separates the three fault modes that matter: bearing wear, cavitation, and misalignment. Diagnostics RAGs the maintenance manuals through Claude; the workflow agent files the ticket. Three MCP servers expose sensor data, knowledge-base search, and ticket management as tools, deployed on Azure App Service with managed Redis and PostgreSQL. [METRIC?: detection accuracy or test count]

Triages inbound complaints end to end: classifies 10 complaint types, reads five sentiment levels from calm to furious, and routes across 7 departments with an urgency tier that auto-escalates critical cases. Replies come back context-aware and bilingual (English and German) in under 10 seconds, against a 15–30-minute manual baseline. Live Gmail IMAP intake, SQLite history, powered by Claude.

Conversational drafting tool where Claude asks the next required question, normalizes free-text answers into structured state via Anthropic tool use, and produces branded PDFs with a legal-review disclaimer. Real bcrypt auth, constant-time verify to defeat enumeration, per-user 'My documents' library, and 142 tests across backend and frontend.

A LangGraph state machine — plan → search → reflect → answer — over a hybrid ChromaDB + BM25 index for German automotive and machinery documentation. The agent chooses from five search tools (hybrid 60/40, pure vector, keyword, and metadata-filtered Excel and PDF lookups) and re-searches up to three times when the reflect node judges results weak. Local all-MiniLM-L6-v2 embeddings, Claude for synthesis; it turns a 30–60-minute manual lookup into an answer in under 10 seconds. [METRIC?: retrieval hit rate / faithfulness]

Production-grade RAG with hybrid dense (BGE) and sparse (BM25) retrieval, RRF, and cross-encoder reranking. It adds an agentic CRAG loop that grades retrieval and decomposes weak queries, three-layer security (prompt-injection guard, content filter, PII redaction), and a reproducible eval harness. The numbers it lands: Hit Rate@5 of 0.96 and Faithfulness of 0.944.

Multi-user trading workstation with real-time Finnhub WebSocket data, simulated $10k portfolios, and a token-streaming Claude assistant that analyzes positions and executes trades from chat.

Pre-market briefing, setup scanner, risk monitor, a plain-English strategy builder with backtests across 54 instruments, security guardrails, a trade journal, and coaching. All of it is orchestrated through Claude, Yahoo Finance, and email.