AI for Traders Has to Fit the Workflow: Moshi's Take

8 min read

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The interview

• You are the Team: Anekha Sokhal, Moshi, on Latent Space with swyx (episode page, show notes and transcript), October 22 2026

Most AI tools for traders will fail, and the model will not be the reason. They will fail because nobody opens them a second time.

That is my take after going through Anekha Sokhal's conversation with swyx, recorded at AI Engineer New York and published on Latent Space on October 22, 2026. Anekha is the CEO of Moshi. She used to trade energy derivatives. Now she is building, in her words, "the program that I wish I had when I was trading."

The interesting part is not how much data Moshi processes. It is how much of the conversation she spends on something else: how a person actually works, and whether your tool fits into that or asks them to change.

One note before we start. The quotes below come from the episode's automated transcript, which Latent Space says may contain errors. I have kept them as published.

Key takeaways

• Moshi says that in the 4 seconds it takes Anekha to say one sentence, it has learned 300 new things.
• Ask it a question and she says you get answers in 30 seconds that would typically take days.
• Her one-line pitch: what you would get if Bloomberg, Claude and Netflix had a baby. The Netflix part, recommendations, is the part that matters.
• Her test for AI tools: a dashboard you build for yourself is "your brain in a dashboard". Getting someone else to use it is a UI problem.
• AI engineers are being told "you are the team" and handed systems that would typically take 10 people to build.
• After the conference she had filled a week of extra meetings, including energy people asking to meet in Houston. Users first, hiring after.

Who is Anekha Sokhal?

Her path explains why she thinks about workflow so much.

She studied mathematics, operational research, statistics and economics at Warwick. She jokes that people called it nerdy back then, and now they hear "machine learning, AI" and get excited. She interned at JP Morgan, started on the emerging markets desk, then moved to commodities and traded derivatives. Most of her career has been in energy derivatives.

She left to become a founder, and she wanted a data company. She got a Fulbright data scholarship to come to the US. She says they gave one in the whole of the UK. Her interview landed in the 2020 to 2021 period, when gas markets had what she calls "insane volatility", so when they asked how she saw the future, she talked about the energy crisis and data. Everyone in the room had felt it in their electricity bills.

In the US she went deep into machine learning research, published at ICLR with Stability AI, and worked in collaboration with NASA. That led to Moshi.

So you have someone who sat in the seat, then learned the machinery. That combination is rare, and it shows in how she talks.

What does Moshi actually do with market data?

Moshi provides real time insights for financial traders, starting with commodity, energy and macro markets. Her short version: "we process more data than is humanly possible to tell you everything that matters to your portfolio right now."

She breaks the inputs down the way she used to break down a market when trading.

Data type	What she listed
Structured, market	Futures time series, technical indicators, moving averages
Structured, fundamentals	Time series for demand and supply in commodity markets
Unstructured, public	News, X data, audio from podcasts
Private to the user	The analysts you follow, your subscriptions, your internal data

That last row is the one I want you to notice. Her example: some energy traders have their own person tracking the number of LNG ships going through the Strait of Hormuz. Moshi cannot tell you who to read. You already know who you need.

Then comes the pipeline. Data cleaning, data transformation, statistics and machine learning, tagging, an ontology on top, and everything goes into what she calls the Moshi "n-dimensional graph".

The output works in two directions:

1. Push. Every time you open Moshi, you see something new related to your portfolio.
2. Pull. You ask, and the LLM gives you an answer in 30 seconds that would typically take days.

Most AI products only build the second one. That is the whole argument of this piece.

Is Moshi a new Bloomberg?

She says no, and her answer is worth reading closely.

swyx, who is a former trader himself, calls Bloomberg the 800 pound gorilla. Anekha gives it credit. She calls the Bloomberg chat "a fantastic aspect of the platform". In the conversation they point out that traders transact billions of dollars in that chat.

But she does not pitch Moshi as a replacement. Her line, which she admits might be controversial:

If Bloomberg and Claude and Netflix had a baby, it would be Moshi.

Bloomberg is the market data. Claude is the interface for working with it. Netflix is the recommendations, tailored to you.

swyx's first reaction was "Netflix?" Fair. Here is why it is the right word.

Why does a trader need a recommendation engine?

Because the thing that hurts you is usually not the thing you are looking at.

Anekha says that if you are looking at one thing, you actually care about the "second and third order derivatives right now, which you wouldn't immediately know." Her example is an oil trader. An energy equities company does something that would indirectly impact oil. You need to know that now.

You're not going to get that from like a news stream. You will get that from a connected graph, which is showing that to you.

She says there are only two ways to catch that link when both pieces of information land at the same moment. A human who connects them, or a system that mimics that human. Moshi built its own recommendation system internally to be the second one.

swyx backed this up from the other side. He runs an LLM RecSys track at every AIE, with people who build recommendations at companies like YouTube, Netflix, Instagram and TikTok. And he says it is not well attended. Engineers do not appreciate how important it is.

I think that gap is the opportunity. Everyone is building the chat box. Very few are building the part that decides what you should see before you ask.

Why do AI dashboards fail inside a real workflow?

This is the section I would send to every developer building an AI tool right now.

Anekha's point, close to word for word:

Everyone now can build their own Claude dashboards. It's like, hey, I built this myself. Great. If you want to get someone else to use it, that's your brain in a dashboard.

Try to get someone new to bring that into their workflow, she says, and you are no longer dealing with a data problem. You are dealing with a UI problem: how do they interact with it?

That is the uncomfortable truth about the current wave of AI tools. Building the thing got cheap. Getting it used did not.

She points at what Spotify and Instagram do well. They push data to you. You do not realise you are on the platform, you just keep finding the thing you wanted. Moshi is trying to work out, as a company, how UI relates to a recommendation system. The episode timestamps even label this part "From dashboards to workflows."

Here is how I read it for anyone shipping AI:

• A dashboard asks the user to come to you, remember what to check, and interpret it.
• A workflow tool shows up where the user already is, with the one thing that matters to their book today.

The first is a demo. The second is a habit.

Isn't a Netflix-style feed just an addiction machine?

This is the strongest counter-argument, and swyx raised it himself. "There's a whole addictiveness element of Netflix, which is bad."

Anekha's reply was two words: "There is." Then she went straight back to recommendation systems.

So here is my answer, and it is mine, not hers.

The Netflix feed is tuned for your time. It wins if you keep watching. A feed for a trader should be tuned for the book. It wins if you catch the energy equities move before it hits your oil position, then close the app and act on it. Same machinery, different target.

The risk is real, though. If a tool like this ever starts measuring success by minutes on screen instead of relevance to the portfolio, it becomes noise with a nice UI. Traders already drown in information. That is the problem Moshi exists to fix, by its founder's own account. A feed that adds to the flood would be worse than no feed.

The second counter-argument is quieter: if anyone can build a Claude dashboard, why buy a product at all? Her own line answers it. Your dashboard is your brain. It works for you. The moment a second person has to use it, you have the UI problem, and you are back to needing someone who has thought hard about workflow.

What does "you are the team" mean for AI engineers?

The title of the episode comes from what Anekha heard from AI engineers at the conference. A lot of them felt overwhelmed by pressure from management.

People now say, we don't need a team, you are the team.

So one engineer is asked to build a whole system that would typically be done by 10 people.

This connects straight back to the workflow argument. If you are the whole team, you cannot afford to build features nobody opens. Every week you spend on the dashboard that only works for your own brain is a week not spent on the part that gets used.

The episode also covers trusted sources and signal, document extraction, and building versus buying models. The show notes link Reducto for the document extraction part. And the episode description sums up her lesson from the conference: meeting users comes before hiring.

She lived that. She was meant to fly out the day of the recording. She extended for meetings on Thursday, then Friday, then Monday, and then people in energy asked to meet in Houston. She had filled a week after the conference with customer meetings.

That is the real workflow research. Not a survey. Sitting with the people whose day your tool has to fit into.

What should you do if you are building an AI tool?

If you take one thing from this, take the push versus pull split. Then do this:

1. Write down your user's day before you write code. What do they open first? What do they check every hour? Your tool has to live inside that, not next to it.
2. Build the push side, not just the chat box. Ask what the user should see the moment they open your tool, without typing anything.
3. Respect their private sources. Anekha's point about analysts and internal data holds everywhere. Your users already know who they trust. Plug into that instead of replacing it.
4. Look for second order links. The valuable insight is often about something adjacent to what the user is watching. That is a graph and recommendation problem, so treat it like one.
5. Measure relevance, not time on screen. If your best outcome is the user leaving fast with the right answer, design for that.
6. Book the customer meetings before the hire. If you are the team, the cheapest way to avoid building the wrong thing is to talk to the people who will use it.

FAQ

What is Moshi?

Moshi is a company building real time insights for financial traders, starting with commodity, energy and macro markets. It combines public market data, news and alternative data with a user's own private sources, then pushes what matters to their portfolio and answers questions in about 30 seconds. Its CEO describes it as what you would get if Bloomberg, Claude and Netflix had a baby.

Who is Anekha Sokhal?

Anekha Sokhal is the CEO of Moshi. She studied mathematics, operational research, statistics and economics at Warwick, traded at JP Morgan on the emerging markets and then commodities desks, and spent most of her career in energy derivatives. She later came to the US on a Fulbright data scholarship and published machine learning research at ICLR with Stability AI.

Is Moshi trying to replace Bloomberg?

Not by her own account. She praises the Bloomberg chat as a fantastic part of the platform and says Moshi is a different type of software. The difference is the recommendation layer, which looks for second and third order effects on a trader's positions.

What does "you are the team" mean?

It is a phrase Anekha heard from AI engineers at AI Engineer New York. Management tells them they do not need a team because the engineer is the team, then asks one person to build a system that would typically take 10 people. Her advice from the episode is to meet users before hiring.

This week, sit with one real user for an hour and write down every tool they open, then build only what fits that list. Grab the free tools on Art of Code for the boring parts, and if you are the team looking for your next role, browse the jobs board.

Written by Adithya Guttha, Founder of Art of Code.