Very delayed news, but back in November @tensorstax was acquired by @Snowflake.
Since then our team has been hard at work building Cortex Code, the current SOTA data agent.
I'll be sharing more of our work/research going forward around harness engineering, cloud sandboxes, eval environments, RL and much more.
Welcome Aria Attar, Biraj Silwal and @tensorstax to @Snowflake !
At Snowflake, we are super excited by how AI and coding agents are completely reimagining data engineering, data pipeline building and the act of working with data.
When we first met Aria and Biraj over dinner a few months back, it was clear we had mind-meld over the future of data and AI. These guys are builders' builders and have great product sense for how to iterate on AI products for data.
Today, we are announcing @tensorstax joining forces with Snowflake. The tooling that the @tensorstax team built to make AI coding agents incredible at data engineering is now live at @Snowflake as part of the Cortex Code product, which launched to GA yesterday.
Read more about @tensorstax joining @Snowflake and what this means for our customers below ...
📷
AI isn’t blocked by models—it’s blocked by brittle data pipelines.
That’s why we’re excited to welcome @TensorStax to Snowflake. Integrated into Cortex Code, agentic systems now reason, verify, and adapt pipelines automatically.
Simplicity scales. Complexity fails. Start building with trusted, integrated AI. bit.ly/3NUWtF2
Transforming natural-language requests into reliable, production-ready data transformations remains challenging. Today, we're excited to announce Thinkquel, our most advanced 32B model for text-to-dbt tasks. Read the full paper below⬇️
By masking out the rest of the response, we get a clearer indication of how much specific wording affected the entropy.
Quantifying context-specific ambiguities yields a sharper view of the data distribution–and a stronger backbone for training better models.
We need to isolate the effect of ambiguity. Thanks to the structured nature of SQL, we can label the exact response tokens corresponding to the minor differences between instructions.
We compare the entropy on the same completion between a vague question and a clear question.
However, entropy is noisy across a response. We need clearer signals for accurate comparison.
Not only is human labeling hard to scale, it's also subjective. How do you quantify how "confusing/ambiguous" an instruction is?
Existing numerical metrics fall short on text-to-SQL. Ambiguity comes not just from the words but also the context of the database schema.
A major barrier to reliable text-to-SQL is that instruction ambiguity is invisible and unmanaged. Our work introduces an entropy-based ambiguity metric, making it possible to systematically evaluate when instructions are underspecified and why LLMs fail.
Full Blog ↓:
We're giving away Coldplay tickets to the first 10 data leaders who book a demo with TensorStax today.
If you're a CTO, VP of Engineering or Head of Data, you're in!
Demo Link Below 👇
Cheat on your Airflow pipelines with TensorStax
Today we're announcing GA of our Astronomer integration.
Whether you're on Astronomer, MWAA, or self-managed Airflow, you can now generate, validate, and monitor DAGs with TensorStax agents, natively.
We're offering 6 months of unlimited seats to Astronomer customers when you activate the integration by September 30, 2025.
Why teams plug in TensorStax:
- Automatically create DAGs and sync to Astro from Jira tickets
- Agents automatically detect issues and fix DAGs in Astro
- Easy integration with dbt, Snowflake, BigQuery, and more
See why the top data teams are using TensorStax alongside their existing data infrastructure: