LH-150
Sessions in LH-150
Who Is the Data Science Consultant? An Analysis of the CNSL Directory
What clustering 949 statistical-consultant profiles reveals about specializations in the profession, and which niches AI will and will not touch.
https://youtu.be/T6l35lcYHwQThe Physics of LLM Inference: From First Principles to Cluster Scale
Why latency spikes the moment concurrent users hit an endpoint, and how continuous batching, PagedAttention, and distributed serving keep throughput high, with live demos of Managed Inference and fine-tune-to-deploy workflows on Crusoe Cloud. We will also dive into Crusoe MemoryAlloy, which reinvents KV caching for cluster-scale inference.
https://youtu.be/qsGOKdJURT4Fixing AI Memory: A Deterministic, Hallucination-Free Architecture
Why LLM-managed memory hallucinates, and a three-tier architecture that keeps the model out of the storage layer entirely, with a live demo.
https://youtu.be/Jeloru6LiBQMy team built a RAG over a product catalog and it was right more than nine times out of ten - close enough to seem right, but wrong enough to kill user trust. The takeaway: if your data has a schema, skip the embeddings and give your agent a structured query tool instead.
https://youtu.be/hSSMkanmjCE?t=6Low Cost/Low Commitment Opportunities for Engaging Students in Data Science
Community college faculty can provide meaningful extracurricular opportunities to strengthen students’ academic interests without significant funding or major time commitments. At Pasadena City College, students participated in the Data Jam and WiDS Datathon, completed independent study projects, conducted research as members of the data science club, and held their first campus datathon. These experiences help students build confidence, develop connections, and see themselves as part of the data science community.
https://youtu.be/hSSMkanmjCE?t=300The Context Window Ate My Steering File
There are ways to minimize the chance that your directive files (eg. CLAUDE.md) will get ignored. Will highlight at least two methods (monitoring the context usage, and the idea of a sacrificial rule)
https://youtu.be/hSSMkanmjCE?t=620The Customer Service Metric We’ve Been Getting Wrong for Decades
For decades, customer service has been built around a seemingly obvious belief: the less effort a customer expends, the better their experience. First Contact Resolution—the king of service metrics—is built on that principle. But after analyzing millions of service interactions, we uncovered a more powerful hidden variable implied by FCR but never truly measured: ownership. Satisfaction barely declines when one person owns an issue across multiple contacts, but it collapses when customers are repeatedly handed off and forced to start over. The finding is profound: intuitive in hindsight, yet disruptive to how service organizations are measured, designed, and operated. So, if we’re being brutally honest, is today’s service architecture actually built to maximize ownership? It isn’t. As AI agents, chatbots, and human experts converge, we must rethink how we measure, route, forecast, and design the future of service around what customers value: continuity, accountability, and ownership
https://youtu.be/hSSMkanmjCE?t=922Everything Everywhere Needs Automation?
Automation is often treated as the ultimate end goal – until it backfires! Join this lightning talk to walk away with two concrete ways to push back on requests for fully automated use cases.
https://youtu.be/hSSMkanmjCE?t=1243Raw Pixels Are a Liability: Building Privacy-First Computer Vision Agents
Raw imagery should be temporary, not the default input to an AI agent. Attendees will learn a practical privacy-first pattern: extract only task-specific signals at the edge, discard identifiable pixels, and let agents reason over minimal structured data.
https://youtu.be/hSSMkanmjCE?t=1461Less Data, Better Insights: Predicting Churn Using Customer Engagement Only
Everyone assumes better churn prediction requires more features. I want the audience to leave questioning that assumption. Customer engagement behavior alone can capture much of the predictive power of a more complex model, showing that more data isn't always better—the right data matters more.
https://youtu.be/Il1qvpr0qwo?t=6Harness Engineering: Data Modeling with AI Coding Agents
When agents hallucinate SQL, the solution isn't better prompts—it's a harness: build a set of deterministic tools the agent must compose from. This overview shows the makeup of a project repo that shows examples of harness engineering.
https://youtu.be/Il1qvpr0qwo?t=323Everything, Everywhere, All at Once: A Better Job for the Robot
Most of the energy going into AI and analytics right now is aimed at getting it to produce numbers you can trust. I think that's the wrong job. In our monthly executive report, the figures always come from code. That single rule changes what AI is actually useful for. Because when a number moves, the hard question was never what changed - it was why. And the answer is scattered across releases, related metrics, past investigations and half-remembered caveats in corners of the business no one person can hold in their head. So I built an agent that reads and connects all of that context, all at once. In five minutes I'll take one real example, walk through the exact chain the agent followed to explain it, and make the case that AI's real job was never the number. It's the why.
https://youtu.be/Il1qvpr0qwo?t=577Practical AI: The Judgment Call Behind a 50% Faster Data Pipeline
Migrating 100+ years of fragmented patient data into a modern Silver layer meant constant pressure to let AI move faster — but not everywhere. I want the room to walk away with the judgment call my team made in real time: which patient-data transformations we trusted to AI, and which ones we never would, because a wrong call there doesn't just break a pipeline — it reaches a patient.
https://youtu.be/Il1qvpr0qwo?t=834Connecting your Agents to Spatial Data
Agents have deep access to text systems — code, documents, calendars — and almost none to physical space. This talk covers why that gap persists, and how one tool closes it. XR Capture is largely solved, but what capture produces is geometry, and a mesh is not something an agent can reason over. Agents need a typed, semantic description of a space — objects with positions, labels, confidence, and stable identity — small enough to fit in a prompt. Companion Scan is an OSS tool that pairs on-device scene understanding with vision models running locally, projecting what the models detect onto geometry the device already holds. Each detection becomes a labelled point with a durable coordinate, anchored so it survives across sessions. That stability is what enables change detection — comparing two scans weeks apart and asking what moved. The session covers the schema decisions that make a space queryable, the failure modes, and letting agents write corrections back.
https://youtu.be/Il1qvpr0qwo?t=1131