AI Hub Build — Instabase
Turning unstructured documents into usable data, without training a model first.
2023–2026 · Product Design, later Head of Design · Design, UX and research
AI Hub Automation Builder
What it is
AI Hub Build is Instabase’s LLM-based platform for building document automation apps. A team uploads sample documents, defines what to extract, tests and validates the results, and deploys the result as a reusable app — one that can classify documents, extract and clean data, and send structured output downstream.
It serves Commercial and Enterprise customers, including in regulated industries, where accuracy, auditability and scale matter.
The problem
Enterprises run on unstructured documents — bank statements, pay slips, W-2s, insurance filings, contracts — but getting structured data out of them used to be slow and specialist work.
Building a solution typically meant annotation, model training, and retraining whenever a format changed. That kept the work in engineering’s hands, made timelines stretch to months, and left business teams with a black box: results they couldn’t easily inspect, test, or correct before they shipped.
The question became: how do you build an LLM extraction flow without training first — and without asking people to trust a black box?
What it could do
Classify documents and split multi-document packets, in plain language
Extract text, tables, lists and checkboxes, and clean the results with prompts, custom code and external integrations
Refine extractions step by step: main prompt → result → refine → validation → final result
Test before shipping, with test sets, result vs. ground truth, and validation pass/fail
Route exceptions to human review, with the source highlighted so a reviewer can verify and correct a value
Deploy as a reusable app, with versioning, deployment integrations and APIs
Early concept sketch for AI Hub Build / Automate
My role
I was engaged on AI Hub Build from the very beginning — designing from scratch, from its early Zero-shot IDP concept to the enterprise-grade product it is today.
Early on, the work was definition and invention: a document-first flow — the document, not the class, is forefront — prompt-driven field definition, and the Extract + Refine framework. Evaluation wasn’t an afterthought; testing and reviewing were built into the product surface, so teams could define what “good” looked like before they deployed.
I'm a named inventor on US Patent 12,067,039 B1, covering the interface for configuring a document extraction flow with a large language model.
As Head of the Design team, I then led the team that made AI Hub enterprise-ready — evolving the Hub, app lifecycle and deployments, and the human review experience — so the same product could hold up for large, regulated customers at scale.