Moshi is a startup founded by former energy derivatives trader Anekha Sokhal that aims to surface real‑time, portfolio‑relevant signals for traders. Speaking at AI Engineer New York with host swyx, Sokhal described her transition from mathematics and trading at JPMorgan to AI research and founding Moshi, and explained how the product works and why it focuses on recommendation workflows rather than being a simple market feed.
Background of the founder
Anekha Sokhal studied mathematics, operational research, statistics and economics at the University of Warwick, then began her career trading on emerging markets and commodity desks at JPMorgan, later specialising in energy derivatives. She relocated to the United States on a Fulbright scholarship, conducted machine learning research (including a publication at ICLR in collaboration with Stability AI) and worked on projects involving NASA before launching Moshi.
What Moshi does
Moshi processes far more data than a single human can read and attempts to translate it into what matters for a trader’s book in real time. The platform ingests and links multiple data types:
- Structured time series: futures markets, technical indicators (moving averages, other statistics).
- Fundamental datasets relevant to commodity markets: supply and demand signals.
- Unstructured public data: news, social media (for example X/Twitter), podcasts and other alternative data.
- User private data: personal analyst sources, paid subscriptions, internal data such as company reports or operational signals (e.g., LNG ship movements).
Sokhal gave a sense of scale: while a human can speak a sentence in roughly four seconds, Moshi typically ingests on the order of 300 new items of information in the same time window. The company builds an ontology and an n‑dimensional graph to connect signals across these sources, and surfaces recommendations and signals tailored to a user’s portfolio each time they open the product. The platform also offers LLM‑driven summaries that aim to deliver in tens of seconds information that would otherwise take days to assemble.
Recommendation system and workflow focus
Sokhal rejects the simple framing of Moshi as a “new Bloomberg.” Instead, she positions Moshi as a different class of product: a blend of market data, language model interfaces, and a recommender that proactively connects second‑ and third‑order effects. Using an analogy she suggested — “if Bloomberg, Claude and Netflix had a baby” — Moshi pairs data access and NLP with a recommendation engine that proposes related items traders may not immediately consider, helping them spot indirect relationships between events and instruments.
The company emphasises that solving user adoption is as much a UI/workflow problem as a modeling problem: pushing the right signals into a trader’s workflow is key to actual usage.
Trust, source divergence and document extraction
Moshi attempts to quantify confidence and context when sources disagree. Divergent reports (for instance, differing counts of shipping activity reported by multiple outlets) are treated as information in themselves: disagreement may indicate an opportunity or uncertainty. Moshi provides confidence scores and surrounding context to help traders interpret such signals.
For messy PDFs and proprietary documents, Moshi uses vision‑language models to parse and ingest user‑provided documents into its database. Sokhal said the company prefers not to forward clients’ sensitive data to third parties; the choice of building in‑house models versus using external vendors (examples mentioned include Reducto) depends on team capabilities and trust requirements. Moshi’s engineering team includes members with research backgrounds, which makes in‑house model development attractive to the company.
Conference feedback, hiring and operational notes
At AI Engineer New York Sokhal reported strong inbound interest: traders, consultants, AI engineers and potential partners reached out after her talk. She said Moshi’s immediate priority is customer conversations and building product‑market fit rather than aggressive hiring, although the company receives substantial recruiter interest—one role posted for four days attracted about 2,500 applicants.
Sokhal also noted a common theme raised by AI engineers at the conference: management pressure that assumes a single engineer can build systems that traditionally require teams (“you are the team”). She highlighted the gap between building prototypes and creating production software that users will actually adopt.
Summary
Moshi aims to be a workflow‑centric tool that filters and connects large volumes of structured and unstructured market information into portfolio‑relevant recommendations for energy and commodity traders. The startup emphasises source trust, user‑specific data integrations, local document processing, and a recommender approach to surface indirect but actionable signals — positioning itself as complementary to, rather than a direct replacement for, existing market terminals.
Organisations and references mentioned
- Moshi (company founded by Anekha Sokhal)
- Anekha Sokhal
- JPMorgan
- Fulbright scholarship
- ICLR, Stability AI (research collaboration/publication)
- NASA (research collaboration)
- Bloomberg, Claude, Netflix (used as analogies)
- Reducto (document processing option)
- AI Engineer New York (conference where the talk took place)
Moshi continues to host customer meetings and webinars to manage incoming interest while iterating on its product and integrations.
