CAS Tech Day 2026

Every new project used to start from scratch. It doesn't anymore. This session walks through the composable building blocks our team has designed to turn new projects from multi-month efforts into multi-day ones — without sacrificing the things that make software real: security, reliability, and automated CI/CD from day one.

We'll cover three blocks in order. First, the foundation: a repeatable serverless AWS architecture that runs at $10–20 per year per solution, is secure by default, and can be dropped into a new project in minutes instead of weeks. Second, the reasoning behind it: the ADRs that shaped the architecture, including the real constraints and pain points that forced our hand, and why the design we landed on is the one that survives contact with production. Third, the accelerator on top: our Claude tools and standards plugin, which encodes our conventions and lets us rapidly develop solutions on top of the foundation without drifting into inconsistency.

Then we'll show it working. A live demo walks through spinning up a new project from the blocks — from empty repo to deployed, production-grade application with full CI/CD pipelines in place — compressing what used to take months into days, with genuine production code rather than a throwaway prototype at the end of it.

When Cloud Services adopted agentic coding assistants, we got the productivity boost everyone promises — and an immediate problem nobody talks about: code reviews became the bottleneck.

That pressure led us to build CRAB, an AI-driven code review bot now running across all ~120 of our repos. But solving one bottleneck revealed a repeatable pattern. We applied the same approach to security compliance — turning policies scattered across four locations into automated, auditable artifacts — and to documentation-as-code, eliminating double-maintenance while giving our AI tools the full context they need.

This talk walks through the journey from that first pain point to a suite of AI-augmented workflows, organized around the three things every team cares about: speed, risk, and cost.  You'll leave with a practical model for finding your own "what comes next" opportunities.

BioCurator is a production-grade, fully on-premises AI curation ecosystem that transforms heterogeneous scientific literature — peer-reviewed manuscripts, patents, and supplementary files across publishers formats — into quality-assured, structured bioactivity data for CAS products including BioFinder. Built on a self-hosted vLLM inference stack, the system orchestrates a modular pipeline with audit-tracked persistence layer, and an extraction engine. A Quality-by-Design normalization layer and a React-based Scientist-in-the-Loop (SiTL) review interface close the curation loop, with every expert correction feeding back as labelled ground truth for iterative prompt refinement. The architecture is now being generalized beyond bioactivity to other domains data extraction, establishing BioCurator as a domain-agnostic, enterprise-scale agentic curation engine operating entirely within CAS data governance boundaries.

Despite rapid adoption of generative AI, most organizations are seeing little real business impact. This session examines why GenAI pilots often fail after early success, where measurable value is actually emerging, and how systems built on trusted tools like Claude deliver results when learning and workflow integration are done right. Designed for engineers and leaders, the talk focuses on making better GenAI bets over the next 12–18 months.

Timely and accurate curation of scientific data by trained scientists is a core tenet of CAS. The products with which our customers access this data are routinely updated to provide them with the best solutions for their work, yet the software used internally by our analysts has not had a major overhaul in nearly 40 years. It is a conglomeration of dozens of disparate programs built over decades in software platforms that are showing their age, and running on hardware that is no longer supported. The development of the Scientific Knowledge Platform aims to bring analyst-facing programs to the same level as customer-facing solutions by creating more streamlined, integrated, and powerful web-based programs. This will bring all curation tasks, from content ingestion to document and structure analysis, to substance registration and reactions into a single unified platform. Additionally, modern AI capabilities are leveraged for a variety of routine tasks, freeing up valuable analyst time for the intellectual value they bring to CAS and our customers.

It's easy to talk about speed and reusability in the abstract. This session makes it concrete by following one project end-to-end. It started as a stakeholder request, was built by remixing our existing building blocks (serverless AWS architecture, Claude tools and standards plugin, one of our frontend options), was shaped through live Tackle Team Tuesday sessions with the business, shipped to production with full CI/CD, and has now been fully handed off to and is owned by the receiving team.

We'll live-demo the app, walk through which blocks were reused versus custom-built, show what "production-ready" actually meant in concrete terms (CI/CD, observability, tests, documentation, on-call), and give a candid account of the handoff itself — what worked, what we'd do differently, and what made the receiving team willing to take ownership instead of sending it back.

What makes Snowflake extraordinary isn't any single feature — it's the whole. This session takes you inside how we've leveraged Snowflake as our unified data platform: from a rock-solid cloud data warehouse that scaled with our business, to governed data sharing and near-zero-copy data collaboration, to powerful analytics that turned raw information into trusted insights. And now, with Snowflake AI layered across it all — Cortex AI, intelligent search, ML-powered pipelines, and generative capabilities — our platform doesn't just hold our data, it understands it. We'll walk through the data assets we've built, the wins we've delivered, and how every layer of Snowflake — storage, compute, collaboration, and AI — is working together to drive our business forward. One platform. No limits.

Optical Structure Recognition (OSR) — the pipeline that detects molecular structure images on patent and journal pages and converts them into InChI and SMILES strings — has advanced rapidly thanks to modern deep learning, yet its end-to-end accuracy is still bounded by two upstream steps: how precisely the detector localizes each structure, and how cleanly the cropped image is passed to the OCR model. This presentation will share two recent improvements to our internal pipeline. First, the structure detection model has been retrained on the latest YOLO architecture, which removes Non-Maximum Suppression (NMS) and Distribution Focal Loss (DFL) in favor of a leaner inference path; the new weights deliver more precise mask segmentation and tighter bounding boxes, improved recall on structures the previous model missed (including starting compounds in reaction schemes and substances embedded in tables), the elimination of thirty or more Markush fragments that were previously captured as false positives, and meaningfully faster inference — with the maintainers reporting up to 43% gains on CPUs and the path now open to deployment on CPU and edge devices. Second, an optional post-processing step based on Maximum Connected Components (MCC) has been introduced between detection and OCR to clean residual artifacts from the mask-cropped image, with a known caveat for tightly-spaced table layouts where the white space between the structure and surrounding ink is too narrow for the algorithm to safely separate; in those cases, the pipeline still produces a usable OCR output that an analyst can finish cleaning. Both changes are intended to reduce the volume of "junk" that subject-matter experts (SMEs) must clean during curation, with at-scale SME validation as the natural next step.

Between us, we've lived three versions of early career engineering: before AI tools, the transition to using them, and onboarding with them from day one. The difference is night and day. Claude Code let us ramp up on complex codebases in weeks instead of months, take on cards we would have avoided, debug infrastructure issues independently, review senior engineers' code with confidence, and turn toil — like parsing a chaotic SharePoint to usable cost data — into real business value.

We also found the trap: when you let AI do the work without doing the learning, speed becomes an illusion. This talk shares the workflows that made us productive, the moment we realized we were leaning in too hard, and the simple habits that keep AI as a tool instead of a crutch. You don't need to be an AI expert to walk away with something you can use Monday morning.

Microsoft 365 Copilot is moving from pilot to broader availability at CAS, and now is the right time to share what we’ve learned. This session will showcase real world M365 Copilot use cases from the last six months. Focusing on how Copilot can accelerate writing, summarize complex information, and support technical and project work without replacing human judgment. Attendees will see practical examples, honest lessons learned, and guidance on how to get value from Copilot as it becomes more widely available.

As AI agents move from experimentation to production, robust observability becomes essential. This presentation explores how Langfuse and AgentCore enable end-to-end observability for agentic systems by providing structured tracing, evaluation, and performance insights. We discuss how these tools help developers understand agent behavior, debug failures, monitor costs and latency, and continuously improve reliability in real-world deployments.

Join Cloud Services as we share the story of our quest to build custom AI agents that perform end-to-end remediation of monitoring alarms. We'll walk through the journey from manually triaging and resolving issues to having AI automation handle a growing majority of alarm response in our Hadoop and OpenStack environments — including the architecture decisions, lessons learned, and what's next.