Startup Fellowship 002: Balto Energy | DevCon 6

transcript

Full verbatim transcript of the Palantir DevCon 6 talk, from a local Whisper transcription of the official YouTube audio. YouTube published no captions. A roughly ten-second pause around 06:33 is a silent demo gap, not omitted speech.

Please welcome Co-Founder and CTO at Balto Energy, Frank Chen.

Hi, I'm Frank. In the last three months, we've entirely rebuilt our company around Foundry and ontologies. Balto is a vertical AI roll-up of home energy companies. People are different, things matter, things change. Every contractor and every home is different. Balto creates digital twins of each. For each time, we use the ontology to make the next digital twin easier and easier.

First, let me tell you where we're going. Balto acquires solar and battery contractors as the wedge into whole home electrification. As an ITC qualified entity, we sell tax offsets to corporate buyers to subsidize deployments, a capital stack that self-funds further acquisitions, each margin accretive. The platform goes from 3.5X to over 8X EBITDA, while digital twins underwrite energy outcomes across a long-lived homeowner relationship that carries every future electrification upgrade.

Next, I'll talk about what we've learned so far. In March, Balto and Northern Pacific signed a term sheet for our first acquisition. In April, Balto joins a fellowship for the first build workshop. The very next week, Balto forward deployed to Northern Pacific for a week-long build workshop. Classic Palantir. Today, we operate as one team. We have over 15 production use cases I know about, each compounding in business value and shaping our trajectory. We have over 500 submissions to our shared company context graph. The magic? Eight key contributors never wrote code before. And of course, this company context graph deployed to Foundry is our operations layer. We are still so early. We have line of sight to make 84% of our first op-co's tasks AI augmentable or automatable.

Before Foundry, a generic marketing campaign took 11 weeks to launch. Today, a campaign launches in less than a day and are highly customized to the person and context. An example is a 30% open rate and 4% action rate campaign. Similarly, before Foundry, an exceptional salesman had a meticulous multi-data silo task performed over 300 times a year. Multiple agent loops today will save this person at least 250 hours. This type of AI enablement should lead to approximately a 50% reduction in cycle time and subsequently working capital needs.

How does this work? One company, one context graph. The thesis is simple. Our entire company can operate like a code base. Every workflow ships two artifacts, the result and the version improvement to the system that produced it. Again, we create two digital twins. This capture happens differently for contractors versus homes, but for both, we capture learnings into the ontology and operationalize the useful signals to make the next easier to build and create wins. For each great contractor, we connect a multitude of source systems, capture decision traces and tacit knowledge through ESM and shape the ontology around the unique ways of working. For each home in our network, we calibrate digital twins to individual usage patterns, bills, weather and equipment, which then unlocks the ability to create an underwriteable asset class.

Let me show you what this looks like. Here's our current comprehensive view for relationships in the ontology. This is useful to orient our team. Now let's zoom into a few use cases. I'll talk through some of the production loops which are currently expert human gated on a single ontology. This data flow is literally the lifeblood of our company.

Let's start with contractor operations. Recall we acquire operating companies at 3.5X EBITDA. We build compounding value when businesses sell to us. We build loops to understand tacit knowledge that lives in people's heads so it doesn't walk out the door. This is the machine we built to close that gap, accruing value to operating companies and because our companies are employee owned, the crew that builds the knowledge owns a piece of the increase in value.

Let's go deeper on context learnings from our internal ontology. The capture loop. When a PM finishes a step in Salesforce, a home visit or a permit, a 30 second prompt fires right there. A lot of this capture happens by voice. It lands right here as a context capture object linked to the job and the person that captured it. This has happened 123 times since May. Here's one of them. This learning gets fed right back into the ops and sales meeting so front line operators can use it in the coming week. We bring this back to the operating ontology and soon we'll use just in time adaptive interventions to change front line behavior.

Now the homeowner. We build a digital twin of each customer's home because we underwrite energy outcomes. The value we promise comes from the interplay of solar, storage and energy using devices in the house. We start with public records. We add in what the customer tells us about their homes and their plans. Then we layer in actual weather for that period. And then we calibrate against reality. What actual usage was and what their bill was. There are about 250 parameters in our physics based modeling and to size that right, we're in thousands of simulations per home to goal seek with building physics in mind. The discipline is matching equipment, weather and usage instead of getting to the right answer for the wrong reasons. The classic failure mode in building modeling.

When the model is locked, we fan it into synthetic twins. About 17,000 equipment scenarios for a single home. We then collapse that down to a single system that fits what this customer actually wants. That's a homeowner's win. A system sized by physics for their home and their goals, not by a sales quote calculator. And here's where it rhymes with field loops. Every calibration run lands in the ontology with its gates, its flags and its exit reasons. When a run misbehaves, we root cause it. And the signature becomes an object in the graph. This one is fired on multiple homes. The system now recognizes it on site.

When we hit something we've never seen, an agent walks a graph, pulls from the building science literature and proposes a diagnosis. A human checks the work and kicks off the next agent for the fix. Same shape twice. The contractor's loop captures judgment. The homeowner loop captures physics.

Next up is scaling. So we have multiple term sheets out to contractors. We are growing our world-class team, raising our first priced round and offering tax benefits to C-Corps. I want to share a special thank you to our fellowship cohort and to the Palantir crew. Thank you so much. Onwards.