Big news at Samaya: our Founder and CEO, @maithra_raghu, has been named to the @TIME 100 AI 2025 list! 🎉 Through her AI research at Google Brain, working with the "Godfather of AI" Geoffrey Hinton, to building @samaya_AI, Maithra has been driven by the vision of Expert AIs that

Live testing GPT-5 on challenging financial questions and still seeing hallucinations and other errors crop up. The automatic routing to thinking is also frustrating at times --- GPT-5 thinks for a long time only to provide a subpar answer. (See below for an example on Figma's

Now generally available. 👉

Under the hood: it draws from reasoning frameworks like Causal World Models 🌍 ( https://t.co/Mvaj60xqQZ) So it doesn’t just summarize headlines, it simulates ripple effects across markets, sectors, and systems.

✅ Analyzes world events using second-order reasoning to surface market-moving insights ✅ Eliminates hallucinations using primary financial sources ✅ Cites exact sentences in the source docs, no vague summaries ✅ Built specifically for financial workflows (not generalist

Most AI “deep research” agents hallucinate 🤖💭 And almost none can connect second-order effects. After 4 months of iteration with users, we’re launching Samaya’s Deep Research Agent, built to solve both problems. (📸 example below ⬇️)

Congratulations to @figma 🎉 Use Samaya to dig into Figma's S-1 and research breaking market updates in real time.

🌍 𝗚𝗹𝗼𝗯𝗮𝗹 𝗖𝗿𝗶𝘀𝗶𝘀 𝗗𝗲𝗺𝗮𝗻𝗱𝘀 𝗖𝗹𝗮𝗿𝗶𝘁𝘆 — 𝗦𝗮𝗺𝗮𝘆𝗮’𝘀 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗗𝗲𝗹𝗶𝘃𝗲𝗿 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀 𝗶𝗻 𝗥𝗲𝗮𝗹 𝗧𝗶𝗺𝗲. As the Israel-Iran conflict unfolds, markets are shifting fast. Our latest agent-generated report breaks down the potential

This brings us closer to our goal: An assistant that doesn’t just respond—it thinks ahead. By detecting and filling gaps automatically, Samaya delivers answers experts can rely on. Full story → https://t.co/9IOnCI9mCA

Impact? 📈 Up to 60% more facts recovered on complex queries 📊 10.5% boost on our toughest benchmark (Criteria-Eval)

Here’s how it works: As Samaya generates an answer, a second process checks for gaps. If anything’s off, it launches targeted background searches—quietly and in real time. It then uses the additional evidence to improve the answer, without noticeable additional delay.

Even great answers can have subtle gaps—missing numbers, overlooked facts. In high-stakes workflows, those gaps cost time and trust. At Samaya, we built a real-time self-correction engine that quality-checks our summaries and fills in missing info before the user notices. 🧵👇

With over 3,000 carefully annotated queries and 8,000+ hours of expert annotations, Criteria-Eval has transformed how we monitor, debug, and continuously improve our platform. Read more details in our blog post:

Why Criteria-Eval - Reproducible: Clear binary criteria enable objective scoring - Flexible: Checklists allow multiple valid answers - Expert-aligned: Directly encodes expectations of experts - End-to-end: Evaluates full pipeline interactions, not isolated components

How Criteria-Eval works 1. Given a query, domain experts (finance) provide an exhaustive list of yes/no criteria each answer must meet, e.g., "Apple’s fiscal 2023 revenue was $383.3 billion". 2. LLM judges score each criterion individually 3. Score = % of criteria satisfied

Evaluating long-form answers to complex technical questions is very challenging. Existing methods fall short in this setting. At Samaya, we built Criteria-Eval, a checklist-based evaluation that aligns with how domain experts judge answers. 🧵 https://t.co/BxidFqA7hc ✍️

🎉Big news: We’re announcing our Series A led by NEA and $43.5m in funding to build the future of expert intelligence in finance. Led by @NEA with support from visionaries like @ericschmidt, @ylecun, David Siegel, and Marty Chavez. Trusted by premier financial institutions like

Everyone's talking about #AgenticAI, but what does it actually look like in finance? @Cortandr from @samaya_AI shares how we build AI agents that domain experts can trust and use every day. Real workflows. Real stakes. 🎧 https://t.co/sV9sfCeN27 #FinTechAI #RAG #AIAgents

🧐 Real-world queries can be long and complex in RAG applications - but today’s retrieval models often fail to understand the subtleties in them. 🚀 In our latest blog, we introduce ✨ Promptriever ✨, our #ICLR2025 work that supercharges retrieval models with instruction