Turing

@EPMgift @DeepLearningAI The best!

@turingcom @DeepLearningAI Great people

Rooftop HH during @DeepLearningAI's AI Dev 26 X SF last week and the conversations went far beyond the sessions. Builders, researchers, and operators sharing what’s actually working in AI right now. Real deployments, real constraints, real impact. If you’re building in AI and want to be part of what’s next, we’d love to connect.

@jeffreyweichsel Welcome!!!👏🏻

May the Fourth be with you! I’ve joined the Light Side at @turingcom as Head of Frontier Data - Enterprise Workflows. Excited to wield the Force of frontier data and defeat the dark side of entropy. Working on synthetic data or complex workflows? Let’s connect.

Paper: arxiv.org/abs/2603.13594

Excited to share that @Turingcom contributed to EnterpriseOps-Gym, a reality-check, top-downloaded enterprise benchmark for agents from @ServiceNowRSRCH and it has been accepted accepted to the upcoming @icmlconf in Seoul, Korea! Stay tuned for EnterpriseOps-Gym v2 — more dimensions, harder setups & richer failure modes. Website & Paper below.

Explore RL environments, evaluation infrastructure, benchmarks, and real-world data systems for post-training research: go.turing.com/turing-at-ai-d…

Couldn’t agree more. Speed will come from systems that run the deployment loop.

@EzgiKorkmazAI @icmlconf Congrats!!🎊

✨My single author paper just got accepted as a spotlight presentation at @icmlconf. Truly excited! More news to come! #ICML2026

Data companies don’t see deployment. Deployment companies don’t see data. Turing sees both.

The best AI agents complete enterprise workflows correctly only 37% of the time. That's not a model problem. It's a benchmarking problem. ServiceNow built EnterpriseOps-Gym to fix it. Here's what they found: Most benchmarks test short, static tasks. Real enterprise work looks nothing like that. @Turingcom designed 1,000+ prompts across 8 domains: HR, ITSM, CSM, Email, Calendar, Drive, Teams, and hybrid workflows. Each task required 7 to 30 sequential steps with real policy constraints. Results from frontier models: → 37.4% task completion at best → Planning, not tool access, is the #1 bottleneck → Human-authored plans boosted performance by 14 to 35 points → Policy compliance was unreliable across the board If your agent hasn't been tested on long-horizon, stateful workflows, you don't know how it actually performs. Full case study below.

Accelerate and innovate your IT projects: turing.com/get-started

Learn more: turing.com/resources/if-a…

Most enterprise AI does not fail because of bad models. It fails because of missing engineering discipline. You would not deploy a pricing engine without version control. You would not release a claims system without regression testing. You would not run a trading platform without monitoring. AI is now making those same decisions and is rarely held to the same standard. Every SDLC discipline maps directly to AI workflows. Unit tests become evaluation suites. Release management becomes model versioning. QA becomes human-in-the-loop validation. Application logging becomes model observability. Performance testing becomes drift detection. Without these controls, the failures are quiet. Models degrade silently for months. Outputs cannot be traced back to a specific version or input. "We did not know the model changed" is not an acceptable answer to a regulator or a legal challenge. The hardest part is not technical. It is ownership. Data owns training. Engineering owns deployment. Compliance owns policy. AI fails in the handoffs between those teams, and that gap stays invisible until a regulatory exam or incident forces it into view. The organizations getting this right have made one shift: they stopped treating AI as a series of experiments and started treating it as infrastructure. The tooling and frameworks already exist. What is missing is the institutional expectation that AI systems should be held to the same standard as the software they sit alongside. Full breakdown below.

Turing built a production-ready RL Gym for training AI agents on real commercial sales workflows. The scale: -100+ workflows -4 platforms: LinkedIn Sales Navigator, HubSpot, Outreach, Calendly -50+ cross-platform tasks -Pass@3 difficulty calibration across all workflows What’s inside: -Sandboxed UI replicas with realistic data -Natural-language prompts tied to structured execution steps -Step-level verifiers with assertion-based scoring -Standardized API for consistent reward signals -Dockerized environments for plug-and-play training Why it matters: Agents don't learn from screenshots. They learn from consequences. This system enables: • Verifiable task completion • Step-level failure analysis • Multi-platform coordination • Scalable RL experimentation This is how you move from demos to real agent performance. Learn more below.

We’re live at @DeepLearningAI’s AI Dev 26 in San Francisco. Come see how Turing is advancing the next wave of AI, from expanding model capabilities to delivering real-world impact at scale. Stop by booth #113 to meet the team and get a closer look at what we’re building. And if you’re around this evening, join us for our rooftop happy hour from 5:00 to 8:00 PM. A great chance to connect, unwind, and keep the conversation going. Details below.