Excited to announce @tensorlake Cloud! 🧵 Tensorlake converts real-world documents into clean, structured data for business workflow automation and for building Agents in mission-critical documents. It's powered by a state-of-the-art document layout understanding model trained

Huge W for humanity if this works out at scale. https://t.co/xwWNr0dLeN

Unless they can deliver sub milliseconds latency this will go no where.

You can now login into Tensorlake using Microsoft and Azure SSO credentials! This is the beginning of better integration with Microsoft Azure and Tensorlake. If you are using Azure, and need better Document Ingestion and ETL for unstructured data reach out to us!

A little secret that everyone building server-less systems know but I don't see being shared often - you have to move away from Docker’s layer-based image format, which relies on stacking tarball layers using overlay-fs, to a file/block based indexing approach. This makes it

if you feel like you're not running enough experiments, the bottleneck is almost always infrastructure, not ideas. focus on improving your infrastructure: write parallelized code: many teams are still doing all their tests using for loops. spending 1-2 hours learning to write

Hard disagree. A/B Tests are cheap and often provide good top line indicators if a new feature improves a top-line business metric. Evals are great and products that use AI shouldn't be shipped before running evals, but there is a place for A/B tests to continuously run in

Really impressed with the simplicity of @trychroma's write ahead log implementation on S3. I hope future cluster schedulers build on top of this to store their state! Most schedulers(Mesos, Nomad, K8s) has been built on top of replicated state machines with state on SSDs and

@diptanu @tensorlake Love this direction. Structured extraction is powerful but without traceability it’s hard to trust. Bounding-box citations close that gap and this will be well received.

@diptanu @tensorlake Much-needed feature, Diptanu. Just tested this on a few complex docs, and it works perfectly. Thanks for releasing this. Really helpful.

How do we trust data from AI Workflows? - We have heard this time and again from developers working in finance and healthcare. Today, we’re releasing citations for Structured Output in @tensorlake Now you can build verifiable AI workflows. Every field extracted has page number

This is a really excellent team (and has some SlateDB contributors on it!).

We need a distributed systems pod with builders. There is a platform shift happening in infrastructure. Early stage companies building on server-less more than ever, databases going from SSD to S3, pretty much every Neo/AI cloud building a container runtime for speeding up

We are hiring a backend engineer at @tensorlake. If you love Rust, and want to work in the intersection of data and distributed systems, this role is for you! DM or email me. More information here -

Coding LLMs seems to write simpler code the more you chat with them about requirements. The first version is always more complex it needs to be or too simple to solve the problem. Experienced developers should be able to understand a solution is too complex and guide the LLM to

We are bumping into a problem where @tensorlake's OCR models perform better than ground truth of OCR Bench v2 benchmark. On some documents we are getting penalized for getting accurate answers 😅 Ex - Ground Truth is missing the subscripts next to "VASP"

This is really cool. Research like this is going to enable parametric memory some day. In the short term external memory retrieval is the way to go. Worth keeping an eye on this line of research.

I wish they had good documentation and tools to go with the hardware.

How does SEA have a Qdoba. Is this the only place in west coast where they have a store?

Like this idea. Once you ship an MVP Agentic application: watch how users actually use it → refactor into deterministic workflows that just work. Most teams are focusing of pre-production benchmarks on synthetic data. We will see more discussions around in-production tracking

Pretty much how I feel when I try to make it write a non trivial feature in a state machine.