Kashikoi Thinks AI Can Finally Make Commodity Trading Less Miserable

The Macro: Commodity Trading Is Stuck in the Paper Age

Physical commodity trading is one of those industries that looks sophisticated from the outside and runs on duct tape from the inside. I am talking about the people who buy and sell actual oil, grain, metals, and natural gas. Not the futures traders staring at Bloomberg terminals. The ones arranging physical shipments, managing letters of credit, verifying certificates of origin, and reconciling invoices across dozens of counterparties.

The document load is staggering. A single physical commodity trade can generate 30 to 50 documents: bills of lading, inspection certificates, quality reports, insurance documents, trade confirmations, shipping instructions, and on and on. Most of these still move through email as PDF attachments. Traders and their operations teams spend hours every day reading, cross-referencing, and manually entering data from these documents into their systems. It is boring, error-prone, and expensive.

The big trading houses like Vitol, Trafigura, and Glencore have internal tech teams that build custom tooling, but even their solutions are patchy. Mid-market trading firms, the ones doing $500 million to $5 billion in annual volume, mostly rely on a mix of spreadsheets, legacy CTRM (commodity trading and risk management) systems, and a lot of human labor. The CTRM market itself is dominated by older players like Openlink (now part of ION Group), Allegro, and Brady Technologies. These systems handle position tracking and risk but do very little to reduce the document processing burden.

There have been attempts to fix this. Fintech companies like Komgo and Vakt tried blockchain-based solutions for trade documentation in commodities. Both found that the industry moves slowly and getting counterparties to adopt new platforms is brutally hard. The blockchain angle turned out to be a solution looking for a problem. What traders actually need is not a shared ledger. They need someone, or something, to read the pile of PDFs faster.

That is where AI fits. The combination of large language models that can parse unstructured documents, agentic workflows that can chain together multi-step processes, and integration layers that can push data into existing systems is a genuinely good match for this problem. The question is who builds it first and whether trading firms will trust it.

The Micro: Enterprise AI Meets Physical Trading Desks

Kashikoi is building an AI copilot specifically for physical commodity traders. The company automates document-intensive workflows using agentic AI, meaning the system does not just extract data from a single document. It chains together multiple steps: reading a bill of lading, cross-referencing it against a trade confirmation, checking the terms against the contract, and flagging discrepancies. That multi-step capability is what separates this from a generic OCR tool or a ChatGPT wrapper.

The founding team brings a mix of AI depth and security research. Aaksha Meghawat led the simulation and evaluation stack at Moveworks (acquired by ServiceNow), where she shipped tooling for 250+ enterprise AI agents. Before that, she built edge speech recognition models that ran on over a billion iPhones, and her research on Transformers at Carnegie Mellon won NSF funding and an Interspeech 2021 Best Paper nomination. That is a strong AI resume, particularly the enterprise agent experience at Moveworks, which is directly relevant to building agentic workflows for trading operations.

Tim Michaud, co-founder, comes from a different world entirely. He is a security researcher with multiple CVEs across Apple products including Safari, macOS, iOS, tvOS, and iTunes. He also discovered a high-impact vulnerability in Qualcomm GPS chips. Security research and commodity trading AI do not have obvious overlap, but the pattern recognition and systems-level thinking that makes someone good at finding exploits translates well to building reliable enterprise software. Trading firms care deeply about data security, and having a co-founder who has literally broken into Apple products might actually be a selling point.

They are part of Y Combinator’s Spring 2025 batch, based in San Francisco and Singapore. The Singapore presence is smart. Singapore is one of the three major commodity trading hubs globally, alongside Geneva and Houston. Being in Singapore means proximity to the Asian commodity trading firms that handle enormous volumes of oil, LNG, palm oil, and metals. Many of these firms are less tech-forward than their European and American counterparts, which could mean both a bigger opportunity and a harder sales cycle.

The product is live at getkashikoi.com, though the marketing site is still sparse. That is typical for a B2B product at this stage. The real product lives behind a login screen and gets sold through demos, not landing pages.

The Verdict

I like this bet. Commodity trading has all the ingredients for a successful vertical AI company: high document volume, expensive human labor doing repetitive cognitive work, clear ROI from automation, and incumbents who are not building this themselves. The mid-market trading firms are the sweet spot. They are big enough to have the pain, small enough to not have internal AI teams, and desperate enough to try a startup’s product.

The risk is sales cycles. Commodity trading firms are not known for fast procurement. Getting past compliance, IT security reviews, and the general conservatism of an industry that still faxes documents in 2026 takes time. The Singapore angle helps because it gives them access to a concentrated cluster of potential customers, but each deal could still take 6 to 12 months to close.

At 30 days, I want to see how many trading firms are in pilot. Not “interested,” not “in conversations.” Actually running documents through the system. At 60 days, the metric that matters is accuracy. Trading documents have real financial consequences, and a single missed clause in a letter of credit can cost hundreds of thousands of dollars. If Kashikoi can demonstrate accuracy rates above 95% on complex multi-document workflows, the product sells itself. At 90 days, I want to see whether they have landed a mid-market firm as a paying customer or whether they are stuck in the pilot-to-paid conversion gap that kills most enterprise AI startups. The problem is real. The team is strong. The question is whether commodity traders will trust an AI to touch their paperwork before their competitors do.