Launch HN: Trellis (YC W24) – AI-powered workflows for unstructured data

Hacker News Discussion ↗

AI startups are targeting the huge volume of PDFs, emails, and scans that still trap critical enterprise data, promising to turn these into reliable structured records for analytics and automation. Commenters see strong demand—especially in finance, logistics, and healthcare—but stress that near‑perfect accuracy, validation, human‑in‑the‑loop review, and integration with existing ETL and ERP systems are essential for real-world adoption. There is debate over whether such workflow platforms have a durable edge as foundational models improve and large vendors expand their own document AI offerings, or whether the real value lies in domain-specific tuning, compliance, and end‑to‑end workflow management rather than the underlying models themselves.

Technology & Approach

  • Trellis uses a mix of fine-tuned LLMs for extraction/validation/parsing and large foundation models for general reasoning, routed via a model-selection architecture.
  • Vision models are used for complex PDFs and scans (nested tables, images), with OCR/parsers relegated to cross-checking.
  • The system supports connectors to diverse data sources and destinations, workflow orchestration (triggering on new data), and both UI and API for technical and non‑technical users.
  • Validation mechanisms include schema-aware checks, custom rules, confidence scores, reference back to source documents, and “LLM-as-a-judge” style evaluation.

Accuracy, Validation & Risk

  • Claimed accuracy is ~95% out of the box, rising to 99%+ with fine-tuning and human-in-the-loop review.
  • Commenters note that in practice many LLM extraction setups hover around ~90% and that undetected errors are the real concern, especially in finance and healthcare.
  • There is strong emphasis that fully automated pipelines are risky for critical tasks (e.g., payments, medical data); human review remains essential.
  • Some ask about concrete metrics and how Trellis detects failures rather than just average performance.

Use Cases Discussed

  • Heavy focus on financial services: private credit documents, bank risk models, invoices, product catalogs, compliance flagging, and email processing.
  • Other domains mentioned: logistics/shipping, genealogy (vital records), audiograms in healthcare, customer calls/chats, and general ETL for RAG pipelines.
  • Table extraction and multi-row outputs from a single document are a recurring need; Trellis recently added a “table mode” for this.

Positioning, Differentiation & Competition

  • Trellis is framed as “ETL for unstructured data” and an end-to-end workflow tool rather than just a structured-output wrapper around an LLM.
  • Differentiation claims versus NER/rule-based systems: robustness to varied, messy formats and complex tasks beyond simple field extraction.
  • Compared with legacy/document-AI tools and other platforms, Trellis stresses configurability, high-accuracy extraction, validation, and workflow integration.
  • Overlap with adjacent YC startups and open source tools is acknowledged; Trellis emphasizes document-heavy workflows, dashboards, and system-of-record integration.

Market & Business Concerns

  • Some argue this is a feature that cloud/LLM vendors will eventually subsume; others counter that workflow complexity, integration, and last‑mile accuracy are nontrivial moats.
  • Several commenters doubt the “80% of important data is unstructured” narrative and note big banks already have strong ETL teams and internal LLM projects, plus strict vendor/compliance barriers.
  • Others see large opportunities in bespoke, high-touch deployments where the main competitor is expensive human labor, not basic OCR.
  • On‑prem and HIPAA support are requested for regulated industries; HIPAA is said to be on the near-term roadmap, on‑prem is seen as harder due to GPU needs.