Connected and live, this ‘living’ model would show management their current state, what comes next, and what to do about it. Operators know their business better than any consultant, so why not give them that view in real-time, and let them make the calls?

In April, 2023 we started Teragonia on Astradis, the AI operating system purpose-built for PE-backed companies. With Astradis, agents do the work, experts supply the judgment, and the client benefits from the finished result.

Today, Astradis runs inside the portfolios of some of the world’s leading private equity firms.

Origin story: from derivatives audits to decision models

I joined EY at 30, later than most of my peers, after a career lecturing in finance. As far as industry, I selected energy deliberately. This one industry holds nearly every business model there is: refineries, petrochemical plants, offshore engineering projects, LNG terminals, pipelines, oilfield services, shipping, fuel retail, and the trading desk. My work took me deep inside every single one.

My first month was anything but glamorous: 180 bank accounts reconciled manually for one of the largest commodity traders in the world. To escape the tedium of bank reconciliations, I volunteered for the derivatives audits that no one wanted, which gave me a front row seat to the institutions that move markets, from Morgan Stanley and J.P. Morgan to Vitol, Trafigura, the super majors and national oil companies, working alongside people who knew considerably more than I did.

From audit, I moved into M&A due diligence, integration, and value creation: more than 100 transactions worth over $100 billion, from founder-owned businesses with single-digit EBITDA to multibillion-dollar cross-border deals, carve-outs, restructurings, and IPOs. With engineers and PhDs, I built economic digital twins of businesses and mega projects, or living replicas of their economic relationships, kept in step with the real thing through its data. One digital twin carried more than 40,000 variables. My rule for all of them: the math doesn’t count until it shows up as cash.

Large language models arrived as we were forming Teragonia, and they made practical a model I had spent two decades assembling one company at a time.

Bringing decision science to private equity

Moving to Houlihan Lokey brought me into private equity, where sponsors placed hundreds of millions of dollars behind companies they expected to scale three to ten times in a compressed hold, without a public company’s people or technology. For years, technology vendors chased the large enterprise and overlooked PE-backed companies, an opportunity for anyone who can speak the language of EBITDA and technology.

Private equity also spreads what works.

Sponsors have capital at stake, push change from the board, and move fast. A playbook that worked at one portfolio company gives the next a head start, with the sponsor connecting the dots and making introductions.

Sai Mali A., Ph.D., our Chief AI Officer, is the most gifted mathematician I know, a decision scientist with the courage to take AI where its limit. Together, we founded Teragonia with the sole purpose of bringing decision science and AI to PE-backed companies.

Why context is the hard part

The numbers inside the gas contract model example could not be easily explained; the reasons lived with the people who built it. ‘Context’ is what gives meaning to the numbers: the industry, the function that owns every number, and how the business operates. Without it, AI is making a guess. In a 2026 dbt Labs benchmark on insurance data, two current models answered only 63% and 51% of the questions correctly from raw tables. Given the company’s definitions, both answered 100% of the questions those definitions covered.¹

PE-backed companies make context hard to assemble. Few can hire a public company’s depth in finance, revenue, and operations, and their operators want to know first thing what needs action every morning. In a roll-up, the data is scattered as well. For example, one company we met runs 37 accounting ERPs, all of which must reconcile to one set of financial statements that hold up in diligence.

Context is half the challenge. The other half is human.

Business rules, such as how bids are won in a closed room of big-box procurement managers, walk out the door with the people who hold them. One animal protein company gave us eight definitions of material margin in a 30-minute meeting. A transportation company debated what “location” meant: depot, warehouse, lease, cost center, parking lot, or maintenance facility?

Storing that meaning is a technology problem. Making it reliable is an art. Automating it is a science.

Asking people how they think, and getting teams to agree on one definition, since two defensible models can report different retention rates from the same data.

Astradis combines all three into four layers: a reconciled data model, approved definitions, ontologies that hold tribal knowledge, and a live knowledge graph tied to the published financials. Together, they comprise the data foundation and context layer.

Our patent-pending Data Modeling Autopilot drafts the model from thousands of tables for our engineers to approve, so the layers take weeks, and each add-on joins within two to four weeks of close. Every answer is checked against the governed definitions before anyone acts on it, so when a lender or buyer asks how a number was made, the answer is already on record.