Document question answering over internal corpora — ingestion, chunking, embeddings and retrieval, with every answer tied to a cited source.

Retrieval that corrects itself: query rewriting, cross-encoder reranking and a coverage check that re-queries when the context does not actually answer the question.

Entity and relationship extraction into a graph, traversed for multi-hop questions that flat vector search cannot answer.

Tool-using agents behind a deterministic orchestrator — a tool registry with per-agent allow-lists, confirmation rules on irreversible actions, and hard termination limits.

Planner and worker agents with bounded delegation depth, chosen only where an orchestrated workflow genuinely could not do the job.

Model Context Protocol servers exposing internal tools with validated JSON schemas, so one integration works across hosts.

Self-hosted models served with vLLM and Ollama — quantization, batch-size and KV-cache tuning, semantic caching, and routing that escalates to a frontier API only when reasoning demands it.

Speech-to-speech pipelines: transcription, intent classification and synthesis inside a conversational latency budget, with a confidence threshold that hands the caller to a human.

Vision-language models applied to documents and images where plain text extraction fails — scans, tables and mixed layouts.

LoRA and QLoRA supervised tuning followed by preference optimization, measured against a held-out set rather than shipped on impression.

Async services behind every model: FastAPI and Django/DRF APIs, SQLAlchemy data models, Celery queues, and webhook ingestion that acknowledges immediately and processes after.

Service boundaries drawn before the code — deterministic orchestration, provider interfaces that make a vendor a configuration value, and event-driven handoffs between components.

Role-based access as data rather than scattered checks, confirmation gates on irreversible actions, audit written before side effects, HMAC-validated webhooks and encrypted PII.

Containerized services with Prometheus metrics, structured logs and tracing on every tool call, packaged for repeatable deployment.

Tracing on every tool call, prompt versioning with rollback, and fixed evaluation sets scored on each change so regressions surface before users find them.