DatAInnovators & Builders

Jo Hjersman has spent over five years at a healthcare XR startup wearing every hat on the data stack, from research data scientist to Head of Data. In this episode, he walks Saket through how he built and scaled a behavioral analytics platform for mental health care, navigated a full cloud migration from Azure to GCP with critical systems kept online throughout, and keeps the whole system running efficiently on a startup budget.

The conversation gets into the real mechanics: how intent-driven data design changes what you can extract from immersive environments, how to handle schema evolution without breaking downstream consumers, and why the data scientist and ML engineer roles are converging toward an ownership model rather than a specialization model.

Topics discussed:

• Designing intent-driven data capture in immersive healthcare environments

• Building a bronze/silver/gold architecture from scratch at a startup

• Managing schema evolution across multi-device data sources

• Planning and executing a full Azure-to-GCP cloud migration with minimal downtime

• Handling HIPAA-adjacent compliance through UUID abstraction at the schema level

• Why data scientists and ML engineers are converging toward a full-stack ownership model

• What bootcamps miss about real-world data engineering at scale

• Evaluating infrastructure trade-offs from both a technical and business cost perspective

Christine Pierce spent over 20 years at Nielsen building and running the data systems behind a $600 billion advertising ecosystem. She now consults on media measurement, AI deployment, and data monetization. Her take on enterprise AI cuts through the noise: most deployments fail not because of the technology, but because of change management and misaligned incentives.

Christine outlines how the shift from panels and surveys to transactional and big data fundamentally changed the measurement challenge, why deduplication across devices remains unsolved, and how organizations can spot the real barriers to AI adoption before they invest.

Topics Discussed:

• Shift from survey-based to transactional data in media measurement

• Deduplication challenges across devices and platforms

• Independent vs. platform-owned measurement in programmatic advertising

• Why enterprise AI deployments fail: change management over technology

• Individual productivity vs. enterprise-level AI transformation

• Using synthetic data and agentic AI for brand research

• Framework for advising companies on data monetization

• Communicating data ROI in business outcomes, not technical outputs

• Agentic models as the next evolution of programmatic advertising

Most organizations have thousands of dashboards and still can't get a simple answer from their data. Francois Lopitaux, SVP of Product Management at ThoughtSpot, argues that the problem was never the data, it was a fundamental misunderstanding of who analytics tools were actually built for. In this episode, Saket and Francois trace the full arc from dashboard factories to agentic BI, and why the shift from self-service analytics to proactive insight delivery is finally within reach.

Francois walks through how ThoughtSpot's semantic layer approach, built years before LLMs arrived, is now the foundation for its agentic product Spotter. Rather than using text-to-SQL and accepting hallucination risk, ThoughtSpot translates natural language into search tokens first and then generates deterministic SQL, preserving consistency and giving business users a way to verify every answer. The conversation goes deep on context engineering, how to enrich a semantic model with business rules and memory, and why the LLM is only as good as the context layer surrounding it.

Topics discussed:

• Why dashboards failed business users from the start

• ThoughtSpot's semantic layer and search token approach

• How agentic BI differs from conversational analytics

• Why text-to-SQL introduces trust problems at scale

• Combining structured, enterprise, and unstructured data sources

• MCP integration for real-time data and automated actions

• Context engineering as the new governance layer

• Automating semantic model enrichment with AI

• The evolution from reactive dashboards to proactive agents

• How data leaders need to rethink their role in an agentic world

What happens when a company runs 15 AI agents across its processes but still cannot measure their impact on the top or bottom line? According to Yorck F. Einhaus, former Global CDO at Liberty Mutual and CDO at Farmers Insurance, that is not an AI problem. It is a data problem, and it is the most common reason enterprise AI programs fail to scale.

Yorck shares how he led Farmers Insurance through a migration from on-prem to Snowflake on AWS, using that transition as a forcing function to settle long-standing disputes between actuaries, underwriters, and product teams over how the same data should be defined. He also unpacks his framework for decision intelligence, which he applies at every level of an organization: determine only what information you truly need to make a decision, and treat everything else as noise.

Topics discussed:

•        Why insurance is inherently a data business

•        Building physical and technological innovation lab environments

•        Using VR to scale claims adjuster training at Farmers

•        Aligning AI strategy directly to business strategy outcomes

•        Data governance as the primary barrier to AI at scale

•        Migrating to Snowflake to enforce data quality upstream

•        AI and multimodal data in claims, including AI-generated fraud detection

•        Shifting from backward-looking claims history to predictive catastrophe modeling

•        Why lateral career moves accelerate long-term advancement

•        The evolving CDO role in an AI-first enterprise

Most companies think turning on ChatGPT Enterprise and running a few lunch-and-learns counts as AI transformation. Michael Domanic, VP and Head of Generative AI Business Strategy at UserTesting and OpenAI's System Builder of the Year, has spent two years proving otherwise inside an 800-person org.

His starting point is a reframe that changes how the whole program runs: there are no more individual contributors. Everyone is managing a team of three, an assistant and a thought partner with PhD-level expertise available all day, and the real leadership skill now is knowing how to direct that team toward what actually moves the business. From there, Michael gets specific on how UserTesting built the enablement infrastructure to make that mental model stick across every function, how they calculate ROI when half the value is genuinely hard to quantify, and why the companies waiting to see how AI plays out are making the mistake they'll most regret in five years.

Topics discussed:

• Reframing every employee as a manager of a three-person AI team

• Anchoring transformation to three business levers to avoid chasing infinite use cases

• Using custom GPT hackathons to surface bottom-up adoption across all functions

• Running a 2-dozen-person cross-functional ambassadors program as an internal force multiplier

• Quarterly top-10 implementation reviews as a before-and-after ROI measurement framework

• Why functional leadership, not function type, determines adoption speed

• Shifting from model selection to purpose-built tooling as the real enterprise differentiator

• Why transformation requires dedicated leadership and can't be a distributed side project

• Honest framing on job displacement: what the data actually supports vs. what is speculation

Most agents fail in production not because the model is bad, but because they forget everything and can't access the right data at the right time. Rowan Trollope, CEO of Redis, has built his entire product strategy around solving exactly those two problems, and in this episode he gets specific about how.

From architecting a semantic layer that sits between your enterprise data and the agent's context window, to building memory systems that handle conflicting user preferences and temporal grounding, Rowan lays out the infrastructure decisions that actually determine whether agents make it out of POC. He also shares a clear-eyed take on the AI bubble, why he'd put money on infrastructure over apps right now, and what the dot-com crash taught him that still holds.

Topics discussed:

• Why pointing an MCP server directly at backend databases breaks agent reliability

• Redis Context Engine: CDC pipeline plus pydantic object models as a semantic layer for agents

• LanCache: prompt-layer caching hitting 95% cache hit rates and 70% LLM cost reductions in production

• Agent Memory Server: using an LLM to extract, vectorize, and resolve conflicting user preferences from raw transcripts

• Contextual grounding: converting relative time and location references into absolutes before storing memories

• Why the agentic infrastructure stack has not yet solidified and what that means for enterprise adoption timelines

• "Provability" as the framework for predicting which job functions agents will automate next

• The shift from specialist roles to "product builders" and what that means for how software teams are structured

• How Redis became the number one data store for agent workloads by market share, and why agents self-select for simpler APIs

• Why infrastructure is the safer bet in an AI bubble, drawing on lessons from the dot-com crash

Fred Gertz completed his PhD in electrical engineering under the inventor of the modern magnetic hard drive, then left academic research to solve a problem that's stumped manufacturers for decades: how to optimize complex operations when you have almost no data. At Collide Technologies, he's applying swarm intelligence to tackle NP-hard scheduling and routing problems that LLMs fail at spectacularly.

His approach comes from an unexpected place. While most AI startups chase massive datasets and GPU clusters, Fred turned to ant colonies. These insects solve complex logistics problems without central coordination, training data, or computing power. Their collective behavior cracks the same mathematical challenges that paralyze manufacturing floors: which routes minimize delivery time, how to assign hundreds of workers to shifting tasks, what machine parameters balance throughput against reliability.

The methodology borrows from operations research and Taguchi's philosophy, which Fred positions against Six Sigma's dominance. Where Six Sigma optimizes for low variation, Taguchi argued customers deserve the best possible product every single time. That shift in thinking leads to different math: instead of reducing standard deviations, you map how every process parameter mathematically connects to business outcomes like profit or quality. The problem? Operations research textbooks are dense enough to intimidate PhD holders. Collide's swarm algorithms make those techniques accessible to companies running on spreadsheets.

Topics discussed:

• Ant colony optimization combining search functions and route optimization to solve scheduling problems in 20 to 30 seconds

• Operations research and Taguchi methods versus Six Sigma's statistical process control approach for manufacturing optimization

• Delivering ROI with spreadsheet data instead of requiring IoT sensors and six month data collection projects

• IQ OQ PQ validation frameworks from pharmaceutical robotics applied to AI model deployment in regulated industries

• Why NP complete problems are better AI targets than tasks humans already perform well

• Agent coordination across 500 enterprise agents as swarm intelligence's next application beyond LLM reasoning models

• Generating structured outputs from API calls without training data or few shot examples

• Rate limiting and context window management for stateful applications like production planning tools

• Manufacturing data environments spanning paper maintenance logs to live vibration sensors in the same facility

• Evaluating AI without numeric metrics when outputs are text based recommendations rather than classifications

Marcel Santilli built GrowthX to $13 million ARR without hiring an AE until three months ago. His approach: 170 paid workshops at $500+ each validated exactly what the market needed before writing a line of platform code. The methodology behind it came from his time as Deepgram's CMO, where AI workflows plus human judgment generated 3,000 continuously improving pages and helped 4x revenue in three months. His framework for AI implementation challenges the tools-first mentality plaguing most data teams.

Topics Discussed

• Three-role framework replacing traditional GTM engineers: process architects, input calibrators, output bar raisers

• Context engineering as delegation test: if an intern can't succeed with your inputs, neither will AI

• Deep research architecture: internal knowledge bases combined with external signal processing before drafting

• Workshop revenue as product validation: charging $500 to teach eliminates assumption risk

• Cohort analysis through narrative transformation: converting raw session data into plain English before pattern analysis

• Planning layer between research and execution: evaluating task requirements against available context

• Post-processing with human bar raisers: calibrating system improvements rather than fixing individual outputs

• Forward-deployed delivery model: solving customer problems directly reveals what to automate

• Output-first engineering: working backwards from deliverable to determine technical requirements

Companies turn on Microsoft Copilot or Glean, then shut them off a month later after discovering sensitive data exposure across their environment. Stephen Gatchell, Partner and Head of AI Strategy at Ortecha, explains why this pattern keeps repeating and what it takes to actually get enterprise AI into production safely. 

Stephen breaks down the real blockers: unstructured data at petabyte scale that organizations never cataloged, duplicate files spreading PII across 15 different locations, and retention policies that exist on paper but never get enforced. He worked with the EDM Associates to formally define what a data product actually is, and explains why most companies assign data owners without ever telling them what their responsibilities are. His framework for moving AI from lab to production starts with cross-functional governance committees and ends with treating AI models as measurable assets with clear ROI criteria.

Topics discussed:

This conversation was recorded while Stephen was Vice President, Data and AI Strategy at BigID. 

Most AI agent demos still look great but fall apart in production. At BigPanda, Alexander Page's team solved this by building systems that internalize user corrections and improve without requiring source data fixes.

The Engineering Director of Applied AI shares with Saket how his team designs production-grade AI agents for IT operations. When a user flags that step seven of a retrieved runbook is outdated, the system internalizes that correction with appropriate weighting and handles conflicts on future retrievals, even when nobody updates the original Confluence page. He argues this capability is becoming a baseline expectation: users accept that AI systems won't be perfect, but they increasingly expect systems to learn when shown the right answer.

Page also breaks down multi-agent architecture decisions. When you have 100 tools, giving them all to one agent degrades tool selection. His team isolates decision-making by domain, spinning up specialized sub-agents at runtime based on user intent. For evals, they focus on tool call sequences rather than final outputs, making it easier to pinpoint where agent chains break down.

Topics discussed: