Corporate AI Chatbot

Highlights

• 01

  Built an advanced RAG pipeline: Qdrant vector store with department-isolated collections, hybrid dense + BM25 search, cross-encoder reranking, HyDE and multi-query expansion, semantic/parent-child chunking and a hallucination checker; bilingual embeddings (nomic-embed-text for English, multilingual-e5-large for Turkish).

• 02

  Designed a multi-agent orchestration layer (LangChain + LangGraph) with a supervisor routing queries to specialized agents (HR, tech, market, fact-checker, self-critique, translation, summarization) including prompt-injection / PII guards.

• 03

  Served local models with Ollama (Qwen, gpt-oss, llama3.2-vision) behind a multi-provider abstraction with OpenAI / Anthropic / Azure fallback, streaming responses, token-budget context fitting and Prometheus metrics.

• 04

  Designed a feedback-driven DPO (Direct Preference Optimization) fine-tuning pipeline: mined high- vs low-rated production responses into prompt/chosen/rejected preference pairs and customized models via Ollama Modelfiles.

• 05

  Built a separate Turkish subtitle/transcription pipeline with Whisper (whisper.cpp, large-v3-turbo) and a rule-based SRT cleaning engine (~60 Turkish correction rules) producing final SRT/ASS subtitles and DOCX transcripts.