Mehul Shah

CEO, Aryn. Ex-GM, Amazon OpenSearch, Lake Formation, and Glue

[Mehul Shah @Mashah08]

Good morning,

Swam 500m.

I made some good coffee today. The kind that when you drink you can tell it’s a miracle.

[Mehul Shah @Mashah08]

Alright my daughter thinks it’s dumb. I’ll try one more:

That mango was props.

[Mehul Shah @Mashah08]

I wondering how lingo gets started. Let me try something —

I cut a really nice sweet and tart mango today, slightly more tart than sweet on balance. There’s not a good adjective for it so Im gonna go out in a limb and make up one:

That mango was spank.

[Mehul Shah @Mashah08]

Here’s an oldie but a goodie:

https://t.co/8j8KbmCx9d

Randy Pausch’s Last Lecture. If you haven’t seen it, then the hour is well worth it. Even if you have, it’s worth revisiting. Each time I discover something that I had previously overlooked, or I have a completely new appreciation for a lesson from a perspective only gained through experience.

This time I noticed his story about football: “When you’re screwing up and no one is saying anything to you anymore that means they gave up.”

His coaches rode him hard. While it was rough when he was going through it, the lessons stayed with him for a lifetime. I too have been fortunate to have coaches, mentors, and family that continue to give me critical feedback.

“Your critics are the ones telling you that they still love you and care.”

And if you know the lecture, this post in and of itself is a head fake.

[Mehul Shah @Mashah08]

You learn something everyday.

I just found out that Credence Clearwater Revival is from El Cerrito, CA!

Mind blown.

[Mehul Shah @Mashah08]

Today, I made pakoras for breakfast. Yummy. These are deep fried potatoes and onions in a chickpea flour batter with chaat masala. Yum.

Some would argue that deep fried anything is great. I can attest that is not true.

We tried deep frying the soya wadi (think rehydrated soya balls) as pakora. Tasted like deep fried carpet balls.

[Mehul Shah @Mashah08]

I’m going back to the old way of using X — to tell people what I had for breakfast and how my workout routine went.

I made myself an oat milk latte from beans that I got from Mad Goat Coffee this past week. Tasted like fig jam.

And oh yeah, 15 laps, though these laps are smaller so 750m.

Join me.

[Mehul Shah @Mashah08]

See more in the blog: https://t.co/fSkjV9T4bI

[Mehul Shah @Mashah08]

https://t.co/7H9ikHgHst

[Mehul Shah @Mashah08]

@jobergum @simonw I have the opposite impression. Everyone seems to think any answers grounded from external data is RAG. So, it’s impossible to talk about new techniques in context engineering because people think they’ve all been done before.

[Mehul Shah @Mashah08]

In the Batman shows and films, the Joker is the most colorful character. Over the years several actors have portrayed him, each bringing their own interpretation and charisma, if you could call it that. Here are the ones I remember:

Caesar Romero,
Jack Nicholson,
Heath Ledger,
Mark Hamill (voice),
Jared Leto,
Joaquin Phoenix

Of these, my favorite is Heath Ledger. His portrayal got under my skin. I was truly terrified when watching him.

Who is your favorite?

[Mehul Shah @Mashah08]

What once took 20 hours in insurance underwriting now takes 10 minutes.

I've spent months talking to underwriters across companies, and they’re all seeing the same thing: hours cutting and pasting from emails, searching Google Maps, and manually extracting data from submissions.

In 2026, frontier models generate text in seconds, yet underwriters are still cutting and pasting.

They chose this career to build relationships, apply judgment to risk, and grow their book.

Instead, they're drowning in document tedium across multiple jobs: hunting files across email and portals, extracting data from PDFs and scanned faxes, calculating risk in spreadsheets, assembling quotes in carrier systems.

Each step requires switching tools, and every handoff introduces the risk of errors.

The breakthrough comes from systems handling all five steps in sequence, hands-off. Turning a 20-hour submission into 10 minutes frees up underwriters to quote while still on the call. They spend their days building relationships and applying judgment.

Much of the paper-heavy work that currently gets outsourced is already being automated, and that trend is accelerating.

Humans will remain essential, but going forward, automation will give them more time to use their unique judgment and experience.

[Mehul Shah @Mashah08]

What once took 20 hours in insurance underwriting now takes 10 minutes.

I've spent months talking to underwriters across companies, and they’re all seeing the same thing: hours cutting and pasting from emails, searching Google Maps, and manually extracting data from submissions.

In 2026, frontier models generate text in seconds, yet underwriters are still cutting and pasting.

They chose this career to build relationships, apply judgment to risk, and grow their book.

Instead, they're drowning in document tedium across multiple jobs: hunting files across email and portals, extracting data from PDFs and scanned faxes, calculating risk in spreadsheets, assembling quotes in carrier systems.

Each step requires switching tools, and every handoff introduces the risk of errors.

The breakthrough comes from systems handling all five steps in sequence, hands-off. Turning a 20-hour submission into 10 minutes frees up underwriters to quote while still on the call. They spend their days building relationships and applying judgment.

Much of the paper-heavy work that currently gets outsourced is already being automated, and that trend is accelerating.

Humans will remain essential, but going forward, automation will give them more time to use their unique judgment and experience.

[Mehul Shah @Mashah08]

Everyone seems to think AI will replace humans, but I think that's a fallacy. AI is accelerating the work humans do, freeing them from the tedious parts so they can focus on high-value judgment.

Think about an underwriter's day. They didn't get into insurance to spend hours searching through PDFs and copying data into spreadsheets. They got into it for the judgment calls. The relationship building. The strategic risk assessment that requires decades of experience.

Over the next decade, I believe every professional will have an AI agent that handles the undifferentiated heavy lifting. The agent sorts documents, extracts data, and surfaces information needed for decisions.

That leaves the high-value work and intuition for the human. No model can replicate judgment that requires understanding the context you can't quantify.

Agents will provide options, and humans will adjust based on the factors that AI might miss.

What work in your role would you eliminate tomorrow if you could?

[Mehul Shah @Mashah08]

For more information, check out the blog post: https://t.co/W0tmLyjbu3

[Mehul Shah @Mashah08]

Friends, customers, and partners — Today, I am excited to share that @ArynInc has been acquired by @glean !

We started Aryn at the cusp of the AI boom, before the launch of ChatGPT. It was a mad time (and still is). While everyone was (and still is) trying to solve AGI, we wrote a simpler mantra on the back of our shirts: “Answer questions from all of your data.” We worked to use AI to help people in enterprises with their unstructured data.

From the first time I met Arvind Jain (Glean’s CEO), it was clear that Glean’s mission and Aryn’s mission were well aligned. Their vision is big — to enable humans to do extraordinary work — and we like to work on things big. He was humble and decisive. We felt that together we could accelerate and create a way for our combined tech to have the biggest impact. So, why wait?

The Aryn ride has been mad fun. I cannot lie; it was also hard, but the people made it all worthwhile. We are grateful for our customers, including Novacore and @DataRobot, who were a guiding light and without whom we’d be floating in a dark abyss. We are grateful for our investors, Factory (Chris Re, Samuel Jackson, Jon Feiber), @8vc (Bhaskar Ghosh, Kevin Chen, Venkatesh Seetharam, Asanka Jayasuriya, and Karen White), Amarjit Gill, and Lip-Bu Tan who saw something in us before the AI boom and still do. We are grateful for all the partners, advisors, employees, friends, and family whose cumulative blood, sweat, and tears got us here.

As we worked to make the future, we gelled into an unforgettable family. We were many things: inspired, nerds, creators, principled, trustworthy, charitable, and goofy.

We are not bowing out; on the contrary, we are on a bigger pitch shooting at a bigger goal. And while we retire the name and logo, we still carry the message on the back of our shirts.

[Mehul Shah @Mashah08]

A friend once told me — if after some time you’re not making progress towards your goal, you only have two options:

1. Change your goal, or
2. Change what you’re doing.

Seems obvious, but most people don’t.

[Mehul Shah @Mashah08]

Ted Lasso is really good.

[Mehul Shah @Mashah08]

More details here: https://t.co/ADFRdY231M

[Mehul Shah @Mashah08]

🎉 Happy Friday folks 💥— A bonus as we finish off February and head toward the vernal equinox, we're releasing two new features in Aryn DocParse: VLM-based vision pipelines for document parsing, and LLM voting for agentic property extraction.

We have seen Vision language models, or VLMs, become increasingly capable at parsing documents. Today we are happy to announce a new option in DocParse that lets you use PaddleOCR-VL 1.5 model for the entire pipeline from segmentation to OCR to table extraction. It’s both fast and super accurate (SOTA on OmniDocBench).

Our agentic property extraction feature allows you to extract structured information from your documents. The feedback from customers is that they want to know how good the quality is, so we added LLM voting. We perform extraction with LLMs from three leading frontier model providers and use the agreement among them to help you assess the quality. The more LLMs that agree, the more likely the value is correct. It’s also a good sanity check of the quality of your schema.

We can’t wait for you to take it for a spin and tell us what you think. More details in the blog👇