Distyl

Where does an engineer’s work begin? For Marilena Batatoudi, the answer goes well beyond the system itself. During her internship at Distyl, Marilena worked closely with clinicians to build a shared understanding of the problem and the broader clinical decision-making process around it. From there, she was trusted to translate that understanding into a concrete technical direction and lead the conversation around what she’d built. Read Marilena’s story: https://lnkd.in/gMKURmD8 #AIEngineering #ForwardDeployedEngineering #EngineeringCareers

Some of the hardest engineering work starts before the technical problem has been clearly defined. For Emily Broadhurst, that’s one reason why each project at Distyl can feel like its own “little startup.” As an AI engineer, she makes major technical decisions across the entire system—including when to reframe the problem itself. In our first Building at Distyl feature, Emily reflects on becoming a deeper systems thinker by finding a path through ambiguity to the intended outcome. Read the blog: https://lnkd.in/g3shA9aP #AIEngineering #ForwardDeployedEngineering #EngineeringCareers

You can spend months collecting production data without knowing whether your AI system is actually improving. When evaluation infrastructure falls behind, you can ship a change that scores better on your evals but performs worse on real traffic. Before you ship that change, test your evaluation infrastructure across three dimensions: How much production traffic do your evaluation categories cover? Do those categories capture the attributes that actually explain performance differences? And do you have enough eval cases in each category to trust the performance numbers? Measuring these dimensions shows whether your evals still reflect production, so you know if you can trust the evidence when deciding what to ship next. As your evals evolve with real-world use, it becomes clear whether your feedback loop is learning from production or simply accumulating data. Read the full post by Distyl’s Ravi Bhandia: https://lnkd.in/gpBw6Msg #EnterpriseAI #AIEvaluation #ProductionAI #AIEngineering

Modern voice agents can follow complex instructions and even coordinate sub-agents, but they remain brittle on understanding the parts of speech that never appear in a transcript: accent, vocal tension, emotional cadence, and other paralinguistic attributes. We built VoicEmu to simulate those variations and turn them into test suites for voice agents. In an internal VoicEmu ablation, introducing a single paralinguistic attribute increased conversation failure rates by 10–19%. Combining several raised failure rates by ~30%. By reproducing the failures that once surfaced only in real-world conversations, VoicEmu can help voice agents better understand and adapt to the millions of people who rely on them. 🔗 Read the full research from Distyl’s Nikhil Gangaram: https://lnkd.in/gmH-Nrb3 #VoiceAI #AIResearch #EnterpriseAI

Our Summer ‘26 Hackathon wraps today! This week in New York and San Francisco, teams across Distyl have been building agents, skills, and products that improve from their own use, without a human in the loop. The rules: start with a cold run, run multiple interactions with production-level data, and show the delta. More than 20 teams are experimenting with the hardest open problems in how AI systems learn. Some of us are tackling real-time voice, turning our eval suite into self-improving training environments that run rollouts and post-train models. Others are solving “agent amnesia” by building a nightly compiler that makes coding agents cheaper, faster, and more context-aware with every run. The bar for every project is a novel approach in self-iterating systems and a path to production. Want to build with us? We’re hiring: https://lnkd.in/dmH2RAYE #EnterpriseAI #AIEngineering #SelfImprovingSystems #SelfImprovingAI #AIAgents

For AI systems in production, every fix comes with a risk: it may solve the problem it targets while degrading behavior somewhere else. Without clear evidence of how the change will behave across real production cases, experts have to work through each fix from scratch. That bespoke validation work quickly becomes a bottleneck as proposed fixes accumulate faster than teams can work through the evidence required to approve each change. That validation burden is the cost of trust. Canary reduces that burden by turning production feedback into a continuous loop of validated improvement, creating a system that gets better over time. Read the full post: https://lnkd.in/gFDxfruP #EnterpriseAI #AIInfrastructure #AISystems

The skill people work hardest to develop often isn’t where they create the most value. The skill that comes most naturally? That’s usually the one. When AI keeps redefining what valuable work looks like, people can lose clarity about how they contribute. At Distyl, we’ve found that people create value primarily through one of four modes: Builders, who make things real; Operators, who create the systems that keep teams moving; Problem Solvers, who cut through ambiguity; and Artists, who see what others don’t believe yet. The four modes depend on each other, because each one creates the conditions for another to go deeper. Read more on our blog: https://lnkd.in/gDH34u3D #DistylAI #EnterpriseAI 

Recently, all of Distyl came together for our first-ever on-site. Throughout the week, teams across Distyl led showcases on our original research, shared what we’re learning from our enterprise deployments, and dug into some of the hardest technical problems we’re working on. That energy carried into long dinners, impromptu karaoke in the office, and plenty of laughter. We ended the week by sailing past the Golden Gate and celebrating our biggest year yet. We’ve tripled in size over the past year. Together, we felt the scale and momentum of that growth. #DistylAI #EnterpriseAI

AI is making every healthcare workflow faster. The member experience still isn’t getting better. Each workflow improves in isolation, leaving gaps across the member journey. Closing those gaps requires AI that carries context across departments and systems, so the member never has to. Approvals, referrals, claims, and scheduling become one continuous experience, not separate steps the member stitches together. Payers and providers, faced with the same fragmentation at different points in the care journey, are already converging on this member-centric redesign. We wrote about what it takes to get there. 🔗 Read more on our blog: https://lnkd.in/eibyiacY #healthcare #AI #healthcaretransformation #memberexperience