# Key Capabilities
- Universal Task Graph (multi‑workflow): Document classification → field extraction → validation → decisioning → generation.
- Multi‑modal ingestion: PDF, TIFF, PNG/JPG; layout‑aware text extraction + OCR plugin.
- Hybrid models: a) LLM prompt‑programming for zero‑shot, b) SFT (supervised fine‑tuning), c) RLHF (PPO/DPO) with reward models trained on business feedback.
- Feedback fusion: pairwise preferences, star‑ratings, edits/corrections, form accept/reject signals, latency & confidence signals.
- Active learning: auto‑triage low‑confidence samples into a labeling queue (Label Studio/Argilla).
- MLOps: experiment tracking, model registry, canary rollouts, champion/challenger evaluation, drift & data quality checks.
- Compliance & audit: tamper‑proof feedback ledger, PII scrubbing, field‑level provenance.