BusinessMarkets, money & strategy
Moshi, founded by an ex-J.P. Morgan trader, pitches a recommendation engine for commodity and energy traders
Moshi, founded by former J.P. Morgan energy-derivatives trader Anekha Sokhal, frames market intelligence for commodity and energy traders as a recommendation problem. It ranks news, social posts, fundamentals and private data by their indirect effects on each user's portfolio.
From the Latent Space podcast episode of October 22, 2026 and cited sources: Moshi combines public market data with each user's private sources. [1] [2]
SOURCES[1] Latent Space: You are the Team — Anekha Sokhal, Moshi latent.space[2] Transcript: The Recommendation Engine for Commodity… rss.beehiiv.com
AT Pulse research deskPublished · Updated · 4 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- Latent Space podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2506.08194 (Jun 2025) [3] date from arXiv ID
- When
- October 22, 2026 podcast episode published
- Covered
- Oct 24 edition
- Evidence
- 9 sources across 9 sites; transcript sourced
- 1Latent Space: You are the Team — Anekha Sokhal, Moshilatent.space
- 2Transcript: The Recommendation Engine for Commodity…rss.beehiiv.com
- 3GIQ (arXiv 2506.08194)arxiv.org
- 4AI Engineer NYC 2026ai.engineer
- 5GitHub PR: Moshi founder profilegithub.com
- 6StartupHub: Moshi says edge lives in a trader's headstartuphub.ai
RisingHow it’s measured
YouTube video · 7.3K views · 33 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Moshi combines public market data with each user's private sources. The public data includes futures prices, technical indicators, supply and demand fundamentals, news, X posts and podcast audio. The private sources include the analysts they follow, their subscriptions and their internal data. The system cleans and tags all of it into its own ontology, then stores it in a graph. A recommendation layer uses that graph to push second- and third-order effects to each portfolio, such as an energy-equities development that indirectly affects oil. [1] [2]
The system treats disagreement between sources as information. Social posts get a confidence score. When major outlets report conflicting figures, the conflict itself is shown as a possible opportunity. Messy client PDFs are processed by a vision-language model built in-house, so client data does not go to outside document-processing vendors. [2]
Sokhal's lesson on adoption: AI engineers she met after her talk said management now expects one person to build systems that once took about 10. They also said tools sent out by email often go unclicked, even by colleagues who asked for them. Moshi is betting that a good push-based recommender also solves this interface problem. [2]
So far the evidence is the founder's own account in a podcast interview, with no benchmarks or named clients.
Read the full assessment
One implication is worth testing: for AI tools inside finance firms, push-based ranking and data-handling terms may matter as much as model quality.
Executive brief
Moshi is a Houston startup founded by Anekha Sokhal, a former J.P. Morgan energy-derivatives trader. It wants traders to treat market intelligence as a recommendation problem, closer to Netflix than to another Bloomberg terminal. In a Latent Space episode recorded at AI Engineer New York, Sokhal says the system "learns" 300 new things every four seconds. It ranks second- and third-order effects against each user's portfolio, using a document model the company built itself. No named customers, benchmarks or independent evaluations are public. The episode's most useful takeaway for practitioners is about adoption: AI engineers told her their tools get built, emailed out and then never clicked.
What changed and event timeline
Founder's vision paper submitted
Sokhal co-authored GIQ, a benchmark showing that vision foundation models struggle with basic 3D geometry. It was later accepted at ICLR 2026 ().
To 14: AI Engineer New York
This was the conference's finance-focused edition, aimed at banking, hedge-fund and fintech builders. Sokhal is listed as a speaker and as Moshi's CEO ().
Talk and early coverage
A third-party profile repository dates Sokhal's talk, "The Investor's Edge Lives in Their Head," to Oct 20, which is after the listed conference dates (). StartupHub published a skeptical write-up ().
Latent Space episode published
Swyx interviews Sokhal about recommendation systems for traders, private-data handling and post-talk demand ().
Capabilities and access
- Product: real-time, portfolio-specific insights, first for commodity, energy and macro traders (transcript 00:05:46).
- Inputs: futures time series and technical indicators, supply/demand fundamentals, news, X posts, podcast audio, and users' own subscriptions, analysts and internal data (00:06:53–00:08:26).
- Output: a feed that pushes new items each time the user opens Moshi, plus question answering that she says returns "in 30 seconds" what would usually take days (00:08:51).
Read the full section
- Product: real-time, portfolio-specific insights, first for commodity, energy and macro traders (transcript 00:05:46).
- Inputs: futures time series and technical indicators, supply/demand fundamentals, news, X posts, podcast audio, and users' own subscriptions, analysts and internal data (00:06:53–00:08:26).
- Output: a feed that pushes new items each time the user opens Moshi, plus question answering that she says returns "in 30 seconds" what would usually take days (00:08:51).
- Access: no public pricing, launch date or demo; contact-only (StartupHub).
- Model versions: none disclosed.
Technical analysis for researchers and developers
What is documented comes only from the founder's description: There is no paper, eval suite or code, so none of this can be reproduced.
Read the full section
What is documented comes only from the founder's description:
- Pipeline: cleaning and transformation, then statistical/ML tagging into a proprietary ontology, then an "n-dimensional graph" that feeds a recommendation layer (00:08:26).
- Cross-asset links: for example, news about an energy-equities company that indirectly affects oil. These come from graph connections, not a single news stream (00:10:17).
- Source disagreement as a signal: posts get a confidence score. When Reuters, Bloomberg and WSJ disagree, the conflict is shown as information in its own right (00:15:46–00:16:29).
- Messy PDFs: handled by an in-house vision-language model, chosen so client data isn't sent to third parties such as Reducto (00:16:49–00:17:43).
- Agent design: accounts differ. One lists four agents plus a manager (GitHub PR); another describes six deterministic agents (StartupHub).
There is no paper, eval suite or code, so none of this can be reproduced.
Claims and evidence
Corroboration is unavailable for every performance claim. Secondary coverage (BigGo) repeats the podcast's figures and adds no new evidence.
Read the full section
| Claim | Evidence |
| 300 new things learned per ~4 seconds | Founder's statement only (transcript 00:06:20) |
| 2.4 billion data points monitored; "up to 35% more efficiency" | Company messaging, as reported by StartupHub |
| ICLR paper | Confirmed: GIQ, ICLR 2026. It is unrelated to trading. |
| 2,500 applicants to one job in four days | Founder anecdote (00:24:08) |
| Mercury Fund is an investor | Supported: Mercury Fund lists Moshi as a portfolio company. No round size is public. |
Corroboration is unavailable for every performance claim. Secondary coverage (BigGo) repeats the podcast's figures and adds no new evidence.
Context and prior work
- Incumbent: Bloomberg added AI news summaries in January 2025 and Document Search & Analysis for Terminal users (Bloomberg).
- Alternative data has history: swyx recalls buying the X firehose at Balyasny for about $300,000 a month and finding it noisy and open to manipulation (00:14:02–00:14:42).
- Recommendation systems: swyx argues they are underrated in AI engineering (00:11:17).
Read the full section
- Incumbent: Bloomberg added AI news summaries in January 2025 and Document Search & Analysis for Terminal users (Bloomberg). Sokhal says Moshi is a "different type of software," not a new Bloomberg (00:09:52).
- Alternative data has history: swyx recalls buying the X firehose at Balyasny for about $300,000 a month and finding it noisy and open to manipulation (00:14:02–00:14:42).
- Recommendation systems: swyx argues they are underrated in AI engineering (00:11:17).
Limitations, safety and contested findings
- No published accuracy, hallucination or latency benchmarks.
- Trusted social sources can be manipulated. Moshi's own answer is confidence scoring, which is not validated.
- Engagement risk: swyx notes the "addictiveness" side of Netflix-style design (00:11:05). Optimizing for engagement is not the same as optimizing for P&L.
Read the full section
- No published accuracy, hallucination or latency benchmarks. StartupHub flags that there are no audits, named clients or public benchmarks, and that data retention and permissioning are undisclosed (StartupHub).
- Trusted social sources can be manipulated. Moshi's own answer is confidence scoring, which is not validated.
- Engagement risk: swyx notes the "addictiveness" side of Netflix-style design (00:11:05). Optimizing for engagement is not the same as optimizing for P&L.
- Product details conflict: the four-agent and six-agent descriptions don't match.
Business and practitioner implications
- Adoption is the bottleneck. "Your brain in a dashboard" rarely transfers to other users. Moshi's thesis is that solving recommendation also solves UI (00:12:13–00:13:03).
- Small teams, big mandates: engineers report being told "you are the team" and asked to build systems that would normally take about 10 people (00:19:12).
- Data sovereignty drives build vs. buy: keep models in-house if you have the engineers; use vendors if you don't (00:17:36–00:18:15).
Read the full section
- Adoption is the bottleneck. "Your brain in a dashboard" rarely transfers to other users. Moshi's thesis is that solving recommendation also solves UI (00:12:13–00:13:03).
- Small teams, big mandates: engineers report being told "you are the team" and asked to build systems that would normally take about 10 people (00:19:12).
- Data sovereignty drives build vs. buy: keep models in-house if you have the engineers; use vendors if you don't (00:17:36–00:18:15).
- For buyers: ask for backtested signal quality and clear data-handling terms before running a pilot.
The source trail.
Sources (9)
Latent Space: You are the Team — Anekha Sokhal, Moshi
latent.spaceTranscript: The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi
published_transcript
rss.beehiiv.com