Blog
Engineering notes, product updates, and research on applied AI from the bem team.
Anthropic filed its S-1 this morning. We pointed bem at all 420 pages and the exhibits: classify, split, extract, enrich, evaluate, shape, send, join, render and parse, plus an ontology, a knowledge graph, and a chatbot on the File System API. Every output is here.
Antonio Bustamante
A working COI verification pipeline on the bem V3 API. Real ACORD 25 certificates in; limit checks, expiration flags, and a live carrier knowledge graph out. About three seconds per certificate, with a trace you can audit.
Everyone selling "sovereign AI" is selling residency or GPUs. Ownership is a loop: your people correct, your model learns, and you keep the asset.
A hands-on guide. Go from a raw invoice PDF to validated, schema-shaped JSON in a single bem API call, with a full trace you can audit. Real request, real response, no parser required.
Every supplier sends a different format. Extract each onto your internal schema and render them all into one template. Shown on two real vendor epoxy datasheets from 3M and Henkel Loctite.
Extract the details from the document you received, reconcile against your records with Enrich, and render a standardized confirmation. Shown on a real public bill of lading.
Extract the terms from a dense agreement, render a clean source-attributed summary, and keep every value traceable. Grounded in DigitalOcean's real credit-agreement amendment.
Every document generator assumes you already have clean data. Render is the one connected to the part that gets it. We ran it on DigitalOcean's real credit-agreement amendment: bem read the 33-page filing, generated a verified facility summary, and flagged a malformed figure in the filing instead of silently correcting it. New in bem today.
Fable would never. Kimi K3 breezed through it.
We ran the whole S-1 through bem, built a knowledge graph, and counted. The sandwiches check out too: 3,300 locations, $4.3B systemwide sales, a 45%-plus Adjusted EBITDA margin, and the Predecessor/Successor accounting from Blackstone's buyout.
The Supreme Court’s birthright-citizenship ruling runs 194 pages: a majority, two concurrences, and three dissents. We turned it into Docket, an open-source legal-research app, to show what you can build on bem. The full opinion, an interactive knowledge graph, and an agent that answers with citations while reading only what it needs.
A knowledge graph and cross-document memory for your documents, built right out of the box. The best of semantic search and a real filesystem for your agents, with no vector store to maintain.
bem turns the whole prospectus into a knowledge graph you can traverse.
The first n-way matcher for unstructured documents. Match many files into one record, with full lineage on every field.
How to use bem to analyze an S-1 filing: from PDF to orbital GPUs.
Extract, Parse, Enrich, Classify, Send. One workflow that handles every document a logistics operator touches in a day.
Native connectors for Google Drive, Dropbox, SharePoint, Box, and S3. Available in every account today.
Every extracted field now includes the exact page coordinates of its source. Build provenance into your product with one flag.
Real ambient systems don't ask. They observe and act.
RAG and agent-native document access are diverging. This guide covers both, then demonstrates each on a 456-page NFL contract with working code.
Today we’re shipping the biggest release in Bem’s history: a redesigned API, first-party SDKs in four languages, a Terraform provider, an MCP server for agent integration, and a CLI. Six releases, one principle — deterministic outcomes from probabilistic systems.
Your governance stack only covers structured data. Here is how to close the gap.
It's been a busy quarter. We've been onboarding and working with more and more teams handling extremely sensitive data. We're excited to tell you more about it.
Models like Claude Mythos will coexist with small, specialized ones. They serve fundamentally different purposes. We're building for the side that keeps regulated industries running.
Confidence scores, dual-model evaluation, and the correction loop that makes extraction reliable at scale
Seven infrastructure failures in seven days. This is what happens when speed is the only metric.
Programmatic access to resources has always been non-negotiable. So why is the industry losing its mind?
Because our customers don't need one, and agents aren't ready *yet* for day-to-day critical business operations.
No, Perplexity did not kill Bloomberg. No, Claude Code didn’t kill ServiceNow. Codex didn’t kill Salesforce.
Forge closes the gap between data extraction and a running operation. Drop your data. Deploy.
We are proud to welcome BuenTrip Ventures as a strategic investor. A partnership rooted in shared values, deep ties to LatAm, and a commitment to building the deterministic infrastructure that powers the future of AI.
Zero public exposure. Fully managed infrastructure. Why Private Link is the no-brainer architecture for enterprise AI.
Full video and image understanding, now as infrastructure
Create an agent with bem that can detect products in planograms, CCTV, social media posts, e-commerce, and more
Stop wasting 80% of your payroll on data entry and start generating real-time intelligence.
Invoices can be gnarly. Here's a production-ready, enterprise workflow to extract their data accurately (and enrich it with your source of truth).
bem is an event-driven, function-based AI platform. That's uncommon, but so is AI in critical production environments.
Measure, manage, and guarantee the performance of AI primitives in production. Now it's possible to build AI systems for unstructured data that you can trust.
Ply revolutionizes RFQ management with bem's AI-powered automation
Order.co achieves seamless procurement automation with bee's AI-powered data transformation
Loop automates data transformation at scale for new clients with bem
In this episode, Antonio sits down with Maxim to cover all things AI innovation, infrastructure hot takes, and Maxim's journey prior to Temporal.