Luis Cerqueira

Engineer.
Founder.
Shipper.

I am an engineer and co-founder of Kaval, where AI agents check their facts before they act. I previously worked at Microsoft and Tern. I design, build, and ship software, from crawlers to iPhone apps.

  • Co-founder, Kaval
  • Previously Microsoft, Tern
  • 3,517 commits in the last 12 months

Selected work

I built these four projects this year. Each project has measured results.

Kaval

Kaval tells AI agents when the facts they rely on have changed.

Kaval watches health insurers' policy sites, detects changes, extracts relevant rules, and points to their exact locations in the source PDFs. An agent calls one endpoint and gets a verdict with a signed receipt.

9,904
policy documents from four insurers, with verified file hashes and about $2 in model calls
32,305/32,305
evidence boxes matched the correct location in a frozen test set
$2K→$120K
the first customer's move from a monthly pilot to an annual contract
CrawlingPDF extractionChange detectionSigned receiptsMCP
usekaval.com ↗
A public insurer policy PDF with Kaval's evidence box drawn around the billing code Q5171
This insurer policy shows Kaval's box around the code it extracted.

Tinta

Tinta is a small face below your Mac's menu bar. It hears its name, speaks, and shows its work.

Tinta detects its name on the Mac and speaks through a realtime voice model. It remembers you across your Mac, WhatsApp, and SMS. It completes browser tasks while a second model reviews actions with significant effects. Each reply records the tools Tinta used, so you can ask what it did.

32/32
wake-word attempts detected, with no false alarms in 256 tests
<1%
of one CPU core used while listening
197
tests across the Mac app and its memory service
On-device speechRealtime voiceMemoryBrowser agents
Tinta
I rebuilt Tinta's face for this page. On a computer, move your pointer to make its eyes follow you.

Sharper clicks

I improved a small vision model's ability to find buttons without additional training.

I tested a zoom-and-refine pass around GUI-G2-3B on ScreenSpot-v2. I also documented the failed methods. Reinforcement learning reduced accuracy by 15 points, and distillation reduced it by 1.5 points. Cleaner labels helped more than additional data.

+3.9 pts
higher accuracy on web interfaces
+2.2 pts
higher accuracy on icons, the hardest category
<1 s
per click in the live demo, compared with 6 seconds before
PyTorchvLLMGPU servingEvals
GitHub ↗Hugging Face ↗
Four ScreenSpot-v2 examples mark missed clicks with red X symbols and refined clicks with green check marks.
These benchmark samples show missed clicks in red and refined clicks in green.

Seen

Seen lets you search anything you have seen on your Mac.

Press ⇧⌘Space and enter a query. Seen captures your screens and reads them on your Mac with Apple Vision. It searches the text with SQLite full-text search. If you use the optional model, only text leaves your Mac. If the model is unavailable, Seen uses local search.

100%
of the work for text recognition and full-text search runs on your Mac
7 days
of search history before automatic deletion
MIT
license, plus privacy and architecture documents
SwiftScreenCaptureKitVision OCRSQLite FTS5
GitHub ↗
Seen's search window showing results for 'Where was that note about the launch date?'
Seen finds a note from earlier in the day. The image uses demo content.

Index

Explore more projects. Select a row to read its details.

    The log

    This chart shows daily commits starting on June 1. It counts commits in my local repositories. I use coding agents, so the counts can be high.

    Before this

    I started as an intern, became a lead engineer, and co-founded Lumaa.

    1. 2024 — now

      Lumaa, Inc.Co-founder · San Francisco

      We raised $850K from Founders Inc, Neo, and Thales Partners. I led two engineers who built Lumaa, a product that matched companies to government RFPs. Nivel compared construction bids and cut price-estimate error from 41.6% to 11.5%. Kaval tracks insurer policies for AI agents.

    2. 2023

      MicrosoftSoftware engineering intern · Seattle

      I built full-stack features across three projects for government customers.

    3. 2022 — 2023

      TernIntern, then lead engineer

      I built the MVP in four weeks and grew the product to more than 2,500 users.

    4. 2021

      MaoniSoftware engineering intern · London

      I built the client onboarding app and cut upload time by about 75%.

    5. 2021 — 2024

      MinervaComputer science

      I left to co-found Lumaa.

    Say hi.

    Working on something hard?

    I would like to hear about it. Tear off a tab to copy my email address.