Aparna Dhinakaran

Announcing Exa Research -- Exa's deep research product. Like OpenAI's Deep Research, we can produce extremely in-depth reports. But unlike OpenAI's, ours is an API -- that means you can programmatically run as many reports as you want and specify what structured format to return the data. Basically, you can now spawn an army of research agents in parallel that won't stop until they find the information you need. Powered by Exa search under the hood. Exa Research scores 94.9% on SimpleQA, the highest of any research API. We do it at the lowest price, by using cheaper models for easier tasks, smarter models for harder tasks. The API's easy to use and you can try it here: https://lnkd.in/gW3em34J We're only just at the beginning of the AI agent era. Much more in the works.

250

12 Comments

Stream processing doesn't need to be complex. For years -- stream processors used to be considered powerful, but extremely complex to implement. To run a simple incremental row count on PostgreSQL, you'd typically need: - Debezium for CDC - Avro for schema registry - Kafka for the event queue - Flink to do the actual aggregation - JDBC Sink to push results back to Postgres …plus hours of digging into Kafka, Flink, and Debezium configs. Watermarks, checkpoints, offsets... Fundamentally though -- there isn't a reason stream processing has to be so complex. A new generation of stream processors is flipping the script — combining powerful capabilities with a radically simpler developer experience. They are breaking the traditional tradeoffs between stream processing (fast but complex) vs. batch processing (easy but slower). Parallelism, internal components, fault tolerance, and schema evolution are all abstracted away from the user — and a single simple interface is built to control all the stream processors mechanism (ingesting data, transforming data, sinking data). The result? Companies are building real-time applications that process billions of records in days, not months. The data world is changing, and the barrier to entry for stream processing is getting dramatically lower. If you are interested in learning more about one of these "newer generation" stream processors — welcome to check out Epsio's docs in this link: https://docs.epsio.io

334

28 Comments

Most labs are sitting on a goldmine of underused reaction data. Few are making the most of it. High-throughput experimentation (HTE) can screen thousands of reaction conditions every month - but in most organizations, that data isn’t feeding back into future optimization efforts. Multitask Bayesian optimization (MTBO) is one way to change that - by learning from past experiments to make future ones smarter. Here’s the idea: Traditional Bayesian optimization treats every new reaction like it’s starting from scratch. But what if we could learn from earlier experiments - even if they involved different substrates or conditions? That’s where MTBO comes in. It draws on prior data to accelerate the optimization of new reactions with similar mechanisms, even when the match isn’t exact. This is particularly relevant when carrying out synthesis with commonly used named reactions, such as Amide couplings and Buchwald-Hartwig couplings. MTBO builds a prior from historical datasets, then learns how the old and new reactions are correlated. That relationship helps guide what to test next - getting to better results, faster. By building systems that learn within reaction classes, we can: → Optimize faster → Run fewer experiments → Get better outcomes with less data This is the future of reaction optimization: not just automation, and not just machine learning - but smart, adaptive platforms that improve over time. At ReactWise, we’re making that future real with multitask Bayesian optimization. If you’ve got years of reaction data sitting in ELNs or spreadsheets, let’s talk.

75

3 Comments

I’m very excited to finally share Agent Browser Use in Unify running on OpenAI’s brand new Computer Use model. We’ve been playing with this model for a while now and have found its much better than any open source project at navigating the web to automate complex research tasks. Since launching in October, the Unify Agent has already answered over 1 million questions for customers, across an impossibly wide range of topics. We found our agent was incredible good at finding, extracting, and summarizing textual data from around the web, but couldn’t interact with sites the way a person does. This is a huge problem, the whole internet is optimized for the way humans interact with it. This meant there were certain workflows that would be easy for any sales rep to do, but impossible for our agent to replicate. With OpenAI’s Computer Use model, Unify’s Agents can now interact with sites exactly how a human would and replicate any research workflow a sales rep would. Instead of relying on pre-built data pipelines or text scrapes, our agent plans research tasks, clicks around the web, and extracts precise answers. For example, a climate tech company can ask: “Did this company add EV charging stations in their parking lot?” and Unify’s Agent will: - Find company HQs - Navigate to Google Maps Street View - Analyze photos of the property for EV chargers - Trigger a personalized outbound sequence This process can run across thousands of companies with no human intervention, automating what was previously hours of manual work. We’ve worked closely with OpenAI since our seed round where we were backed by the OpenAI Startup Fund. Now with o1, o3-mini, gpt-4o, and Computer Use, we’re building the first AI system-of-action for revenue teams to help the fastest-growing companies rethink how they engage buyers!

128

20 Comments

Check out our latest blog post on the recent improvements to Helix’s System 1 visuomotor policy. One of the very few examples in robotics so far of an end-to-end AI policy + useful work. https://lnkd.in/g4-i4Uis

58

6 Comments

Prompts now sit at the heart of GenAI systems, moving through a flow that demands real engineering control. The process begins with creating and structuring prompts, adding parameters, configurations, and metadata. It continues with version control, where changes are tracked and teams collaborate. Automated evaluations, metrics, and quality gates then validate performance and reliability. Finally, prompts are integrated into production pipelines, connected to CI/CD, experiments, and monitoring. This flow is not static. Each stage feeds back into the next, enabling constant refinement and higher confidence in the system. Prompt management at this level requires engineering rigor, not one-off creativity. It is the foundation for scaling GenAI safely and effectively. What steps have you taken to make prompt management part of your engineering practice?

93

4 Comments

We teamed up with The Product Folks on mosAIc 2025 - a virtual build a thon for AI builders. As part of the buildathon, there are several hands-on workshop that covers everything from evals, prompting, and agent design to security, infra, and shipping production-ready flows. This is what it actually takes to build a working agentic application. Alongside speakers such as Arvind, Ji Huang, Madhu, Akash, Rohit Khanna, Rahul, Vaibhavi, Eashan, I'll be doing a hands-on workshop as well, to answer questions like: → What truly goes into building an AI agent that can do real meaningful job - layers involved in building it → How do agents work securely on behalf of users and connect with external tools to get the job done. → How should builders think about auth where agents are the primary actors If you’re working on AI infra or products, this is probably the most useful 2 weeks you’ll spend this year. Last few spots left. Link in the comments.

70

5 Comments

We crushed baseline AI accuracy by 7x in our latest evaluations. Most enterprise AI pipelines fail when it comes to complex financial analysis. We've seen this repeatedly with our customers. So we built an open-source evaluation specifically for multi-hop reasoning, time-series analysis, and complex document workflows. The results? Baseline (OpenAI file search): 13.3% accuracy Custom LangChain + SOTA OCR and chunking: 66.67% accuracy Morphik (YC X25): 95.56% accuracy That's a 7x improvement over baseline and 40% better than custom solutions. Some real-world use cases where this matters: - Investment firms analyzing 10-K filings - Insurance companies processing claims with complex medical documentation - Legal teams extracting contract terms from multi-page agreements The difference between 67% and 95% accuracy? A system you can actually trust with business-critical decisions! We have benefited immensely from open-source, so we made the evaluation open-source as well! Happy to take contributions to add your results on there. Check it out: https://lnkd.in/dFp_BErw Building something that needs bulletproof document understanding? Let's chat: founders@morphik.ai. Try it out: morphik.ai

57

4 Comments

Now live: Sutra models via Portkey’s AI Gateway SUTRA-R0 excels on the Massive Multilingual Language Understanding (MMLU) benchmark in languages where high-quality benchmarks are rare, setting a new standard for multilingual reasoning. You can now integrate with Sutra via Portkey’s AI Gateway and get: - Smart routing, fallbacks, and retries for reliability - Guardrails to ensure safety and output quality across languages - Full observability with structured logs, latency, and cost tracking

67

4 Comments

Agentic workflows in enterprise AI aren't just about chaining LLMs— the reality of building effective AI agents involves solving a three-axis optimization puzzle between speed, quality, and cost that defines successful implementation. Increase reasoning depth for quality? Your agent slows down and burns expensive compute.  Rush the workflow to save time? Quality inevitably suffers.  Chase perfect answers? You'll blow past SLAs and budget constraints. This isn't theoretical—it's the reality playing out in every serious enterprise AI deployment: ➡️ Domain-specific chain-of-thought has emerged as the new runbook, making AI agents predictable and efficient where raw LLM chains fail. ➡️ Enterprise data sprawl silently kills AI initiatives. Without proper filtering and structured data extraction, agents spend valuable GPU cycles wading through noise. ➡️ Resource-awareness is becoming the true differentiator. The next wave of AI agents will be judged not just on intelligence but also on how intelligently they consume computational resources. I've explored these critical challenges and our approach to solving them at NeuBird.ai in my latest post. For those building mission-critical AI systems, understanding these tradeoffs isn't optional—it's essential. https://lnkd.in/gjCWrUPZ #EnterpriseAI #AIAgents #ITOps

45

1 Comment

We are excited to launch Argon Intelligence - a mixture of AI Agents specialized in completing end-to-end CI workflows AI Agents enable CI to move from a reporting function to a forward-looking strategic function Argon orchestrates thousands of AIs in minutes to help automate the drudgery of CI. From tracking updates, to drafting alerts and implications, to automatically maintaining indication landscapes We've been testing our CI solution with beta partners and the results have been incredible. Teams no longer have to make tradeoffs between cost, comprehensiveness, speed, or quality

75

6 Comments

Excited to share that I've co-founded a new thing - Neurometric AI, which focuses on optimization and TCO for systems level AI implementations. In particular we use have been building expertise in GPU alternatives. If you don't know what those are, we wrote a guide below. But yes, there are chips out there that have better price, performance, and power tradeoffs than GPUs. Soon we will release our independent benchmarking tests on these platforms. If you need help with that, please reach out.

125

7 Comments