Jason Kealey

Open to trying it out!

Vanna is good for the loading up the schema plus docs into a vector db for RAG part, just the charts part are weaker.

I want the graph part too, that’s where Vanna w/ plotly is failing. It’s not rendering when I use booleans or timestamps and it does scatter plots at inappropriate times

Vanna.AI - Personalized AI SQL Agent

Anyone got open source alternatives to vanna.ai for text to SQL?

The SQL part is pretty good but the plotly charts it recommends are wonky.

I’ve also been playing with dbt - and considered sqlmesh instead.

Chose to go deeper with dbt as I feel like sqlmesh’s real value shines when you want to avoid transforming the data both in dev and prod.

… and I’m just prototyping stuff locally to play with different open source BI frontends.

Last week I did some experimentation with unstructured.io to extract content out of some pdfs.

Also played with github.com/getomni-ai/z... as a more lightweight option.

Overall, vision models do a much better job than classic OCR (ex: tesseract) on tables in docs.

Review of Data Orchestration Landscape ... and honest one ...

Here’s a decent overview of the data pipeline orchestration tools on the market.

dataengineeringcentral.substack.com/p/review-of-...

Does this API exist in both the hosted and self-hosted versions?

When reading the docs last week I sometimes got mixed up in what features needed a subscription.

Nice! Hello!

Is the api giving you just the metadata of the metric or also translating to the SQL you’d run to build charts like you do in lightdash itself?

Lightdash may be a good UI option for me, but then I’m defining metrics at the presentation layer and tightly coupled with it.

I’m looking for something to be able to express KPIs centrally and cleanly, and have the BI layer autogenerated from it.

Dbt semantic layer could be that, but doesn’t seem like many open source BI layers support it.

I spent some time this week playing with some tools to setup a data pipeline for simple BI and data science.

Played with airbyte, dbt, duckdb and metabase.

Planning on trying lightdash next week.

Trying to avoid hosted data warehouses and use open source for the full chain.

Keep it up!

I’ve never felt motivated by hyping whatever I’m building to peers who’d never be clients/users.

I sold my solution by ignoring the sales rules and just flat out asking “how do you do <process>?” and replying we had an app for that when they outlined their manual process.

I felt dirty not outlining benefits at first, but I later realized this opener was better aligned with my ICP (operations)

My past experiment with a tax form was bad because the LLM used basic OCR instead of something fancier for tables.

And here I’m trying to do something more generic without knowing the form format ahead of time.

Imagine a government form with some weird tabular layout to shove as many fields into a condensed space as possible.

Generalized use case is read/write to forms. Basically reverse engineering a domain model from a form.

Azure DocIntel lets you do it well for a known form via training.

If making zip files named _final-website(2)-final-tuesday-3.zip is too complicated maybe something like dropbox with revision history baked in could work - at least for one file at a time.

That’s ironic but I guess totally expected since it’s their main revenue source hah

Interesting compensation model.

#chordle sounded kinda sus

#chloedle

New audio models from OpenAI, but how much can we rely on them? OpenAI announced several new audio-related API features today, for both text-to-speech and speech-to-text. They’re very promising new models, but they appear to suffer from the ever-present risk of ac...

I published some notes on OpenAI's new text-to-speech and speech-to-text models. They're promising, but like other LLM-driven multi-modal models they appear to suffer from the prompt-injection-adjacent problem of mixing instructions and data in the same token stream
simonwillison.net/2025/Mar/20/...

I like hearing “I used X library to do Y and it worked/failed”.

Helps broaden my perspective about what’s out there.

I dislike hearing “do like me and bathe in virgin blood to succeed”.

I don’t know. I’ve been deep in home renos last couple weeks. Haven’t been checking here much!

You’re right, she seems gone!

It’s sometimes easier to get something extra that doesn’t cost the founders more in the short term, such as an extra week of vacation or extra stock options.

My main lesson learned about economics is everyone is mostly wrong when predicting the future even if they have very logical reasons for their predictions.

Because predicting the future is hard.

But you still learn about the motivating factors behind people’s actions in the present.

Love that #econsky is here.

A few years ago, I started listening to podcasts about the economy during my runs and I feel like it massively helped me learn more about how the world works.

I should get into the penetration testing business because there will be so many security holes to find these coming years.

C’est un peu inquiétant de voir qu’avec le climat géopolitique qui se réchauffe que le SCRS rémunère ses agents la moitié ou moins que des postes au privé www.canada.ca/fr/service-r...