Virtual Datasheet

Virtual Datasheet turns PDF datasheets into cited, queryable data for engineers and AI agents: every parameter and graph, each value linked to its page.

  • 748 Raised
  • 457 Views
  • 3 Judges

Tags

  • Tools and Features

Categories

  • Data Handling
  • Room 4
  • Finalist

Gallery

Description

THE PROBLEM

  • Engineers waste about $33 billion a year reading PDF datasheets.*
  • 3 in 4 hardware engineers lose at least half a day every week to datasheets.**

The problem: $33B a year, 3 in 4 engineers

WHY NOW

The datasheet never went digital: printed data books, then PDF in 1993, still PDF today. AI agents now read datasheets, but datasheets are PDFs made for humans, and agents need data.

VIRTUAL DATASHEET

Living Datasheets that answer back. Data for humans and agents.

SOLUTION

  • App: point at any graph and get its exact numbers in seconds, each with its source page.
  • Engine: a vector database engine that turns datasheet PDFs into searchable, cited data.

▶ Demo video: https://youtu.be/kGlPsOt9bjw

A datasheet graph turned into numbers

EXAMPLE: a transistor's datasheet says 49 Amps, but on a real board with no heatsink it handles about 8 Aamps, less than the 9 Aamps our fan needs, so Virtual Datasheet flags it before you build.

Datasheet 49 A vs real board about 8 A: flagged

Ask a question in plain words and get an answer that links to the exact datasheet page.

Chat answer with citations to the datasheet

ARCHITECTURE AND RESULTS (measured at the Lab)

  • 253 datasheets processed.
  • 2,978 cited statements · 218 parts · 2,748 curves digitized, each value linked to its page and position in the original PDF.
  • 90% of test searches find the right answer in the top 5 (21 engineering questions).
  • About $0.16 median extraction cost per PDF (measured on 63 PDFs).
  • 453 automated tests on the extraction and the API.
  • Source systems: manufacturers' PDF datasheets (unstructured documents) and KiCad design files.
  • Data domains: components and manufacturers; electrical, thermal and maximum ratings with their test conditions; graphs; pinouts; design notes; and the source (page and position) of every value.
  • Flow: PDF → extraction (rules + Grok vision) → parts and statements with sources → embeddings (PostgreSQL + pgvector) → engine (Part Matcher, parameter and semantic search) → API (FastAPI) → web app for engineers and MCP server for AI agents.

PDF to Virtual Datasheet to web app and agent-ready API

TECHNOLOGY

  • Extraction: acrescenta que os statements usam o grok-4.5, porque "LLM only as a fallback" não cobre isso. Sugestão: "...An LLM (Grok, xAI) is a fallback for the hard parameters, and it also reads page text and images to write the cited statements. Values it returns must be found on the page or are flagged for review." Confirma com o Miguel a parte "word for word", porque eu não a verifiquei.
  • Provenance: "every value carries SHA-256, page, bounding box" é demais. As 266 peças do VirtualDatasheet citam só a página, e só 63 dos 656 documentos na base têm hash. Sugestão: "every value carries its page, and where available a bounding box and the PDF's SHA-256 hash, so it can be traced back to the original PDF."
  • Embeddings: troca a frase do recall (era 147 statements, hoje são 2 978) por: "2,978 statements across 67 part families are searchable (HNSW index, cosine distance)." A frase do hash ("only new or changed text is re-embedded") não a confirmei, por isso confirma ou tira.
  • Decision engine: "A missing value is reported as unknown, never as a pass" tem exceções conhecidas: o LMR16006 Vout passa sem máximo publicado, e o Id tabelado a 25 °C conta como pass. Sugestão: "A value read from a curve is reported as estimated, and a missing value as unknown, never as a guaranteed pass."
  • AI chat: "the model cannot invent a value without a source" é forte. Sugestão: "Numbers in answers come from retrieved, cited statements; anything else is labelled as general knowledge."
  • Built with: Cursor (AI coding agents opening pull requests), Grok, Python, FastAPI, PostgreSQL + pgvector, OpenCV, poppler and PyMuPDF; frontend in HTML with PDF.js, D3 and three.js.

AMBITION

Not another tool. One source of truth: the data layer between chipmakers, engineers and their AI agents. Every value verified, every value cited.

OUR PLAN

Open source first, to grow the user base. Then first customers. Then a B2B offer for big companies and manufacturers.

WHY US

MAKE IT has designed electronics for clients since 2019: dozens of hardware projects, all built on datasheets. We feel this pain every week.

BUILT AT THE LAB (26 to 27 October)

  • Brought: the problem, from our own hardware design work at MAKE IT.
  • Built here: the web app, the vector database engine, 253 datasheets processed, a real KiCad project and a working demo. Backend repository created on Monday at 12:20, frontend at 22:04; 44 commits in two days.

Hackathon timeline, Monday 26 October 2026

SOURCE CODE

ASK

Help us promote open-source Virtual Datasheet in the SpaceX.ai community. SpaceX hardware engineers spend too much time reading PDFs too.

Plan and ask


* MAKE IT estimate: hardware engineers × hours on datasheets × loaded hourly cost; range $10–91B.

** Accuris / Fuld & Co. survey, n = 439, 2026, vendor-commissioned.

Technical details for judges: Architecture and Data Model (PDF, attached).

Attachments