The CIO Who Built His Quant Desk in Under an Hour

Udayan Mitra had tried AI. As CIO of a $2.6B Scottsdale RIA with 34 advisors, he had used ChatGPT, Claude, Gemini, and Bloomberg's “Ask B” for market research, PPM review, rolling correlations, due diligence on dubious proposals, and weekly market briefs for his advisor fleet. He was not impressed.

“Statistical collation, not intelligence”

Udayan's critique was epistemic, not operational. Generalist AI, he argued, doesn't reason. It mirrors:

He gave three concrete examples:

The World Cup problem

Before the tournament, AI hedged on who would win. After Messi scored three goals in early matches, the AI pivoted entirely: Argentina will win. “It's just scraping from all the stuff people are posting on the web, collating that. But there's really no real nuanced analysis of how long this is going to endure, is one an aberration.” The AI was a mirror of media sentiment, not an analytical engine.

The watch insurance problem

He photographed a watch for insurance pricing. The AI quoted roughly 2x purchase price — defensible as replacement cost, but it never asked the critical question: “How much are you willing to spend for insurance? If you want to insure something for twice your purchase price, the premium is going to be twice what you should be paying.” The AI solved the surface problem without understanding the tradeoff.

The SpaceX IPO problem

Pre-IPO: wildly bullish. Post-IPO, as the media narrative shifted to “100x revenue is ridiculous,” the AI's tone shifted with it. “The tone is changing simply because there's so much media interaction and everything that has been posted that it's just picking up from that.” The AI wasn't analyzing SpaceX. It was mirroring the media cycle.

For an institutional investor whose edge depends on thinking differently from the crowd, a system that converges on the crowd's opinion is useless — or actively dangerous.

The three failures

Beyond the epistemic problem, Udayan identified three specific failures that made generalist AI unusable for his actual workflow:

Udayan wasn't actively looking for a new AI platform. He had just returned from a week of conferences and was, in his words, “conferenced out.” His colleague Ben Jacobson called to catch up, heard Udayan's skepticism about AI, and suggested a conversation about AdvisorClaw. Udayan agreed, mostly out of professional courtesy.

He approached it with skepticism. He'd seen enough AI demos to know that most of them overpromised. His guard was up.

But when Ben Jacobson opened the call not with a pitch but with a question — “What frustrates you about AI today?” — Udayan started talking. And didn't stop.

The GARCH moment

When Ben pressed — “If it wasn't a generalist, if it had memory, if it could learn the way you like to do things, what else would you have it do?” — Udayan went deep:

Then he described exactly what he wanted:

He tied it together with Black-Litterman — the optimization model that takes market weights and adds a “views matrix.” If you do this sequentially, the system learns how your views evolve. It becomes proprietary to you. Generalist AI can't do that.

The ask

Ben didn't sell. He validated: “This is the great financial crisis — take a bunch of bad loans, put them all together, and it won't be so bad if you average them all out. Being a generalist is nowhere near good enough. What you're describing is exactly what we built.” Then he anchored the solution in one line: “Two to three graduate-level employees that learn from you, get better each month, for the compute cost of lunch.”

Udayan didn't need a next step proposed to him. He asked:

28 minutes. No deck. No feature list. A CIO who was “conferenced out” asking to drive in for an in-person demo. He wanted to see it with his own eyes.

The demo took place on July 14, 2026. Ben McMillan joined Ben Jacobson for a 54-minute Teams session with Udayan. McMillan opened by framing the philosophy:

Then he got to work. What follows is the workflow McMillan walked Udayan through — the same one captured in AdvisorClaw's “Building a Volatility Model” training video.

Step 1 — Upload the data (00:00–03:20)

McMillan had Udayan send over his Apex Equity Growth strategy returns — data he hadn't seen before the call. He uploaded it to a fresh agent and asked it to take a look. The agent spent roughly a minute accessing the file, then responded:

McMillan hadn't known it was monthly, or that there were two streams. The agent figured it out. Udayan was watching.

Step 2 — Verify understanding (03:20–03:48)

Before running the model, McMillan paused to verify Udayan's comfort with the methodology:

The agent produced a clear explanation of GARCH(1,1) — the autoregressive and moving-average components, why it's preferred over naive historical volatility, and the Nobel Prize–winning lineage (Engle, 1980s, now at NYU). Udayan, an econometrician who applies advanced statistical models to financial data, saw the system pass a basic competence check.

Step 3 — In-sample calibration (03:48–06:06)

The agent calibrated GARCH(1,1) in-sample on both return streams, using Python with pandas — “because we tell it that's what you want to use anytime you're using quant finance,” McMillan noted — and produced GARCH omega, alpha, and beta parameters for each stream, the half-life of volatility shocks, and long-run (unconditional) volatility estimates. Then it interpreted the results unprompted:

Udayan's strategy is an actively managed stock portfolio. The agent was identifying characteristics of managed returns — low skew, serial dependence — from the data alone. McMillan, who's used GARCH for 20 years, confirmed the agent was producing PhD-grad-student-level output.

Step 4 — Walk-forward validation (06:06–07:07)

McMillan didn't stop at in-sample. He wanted out-of-sample validation — the standard a trained statistician like Udayan would demand. The agent proposed a walk-forward plan: calibrate on the first 100 observations (~10 years of monthly data), validate on the remaining 50, then expand the window. This is how you test whether the model actually works out of sample.

The agent went further than asked, running five model variants instead of the two requested: standard GARCH(1,1), EGARCH (captures leverage effects), and GJR-GARCH (treats upside vs. downside volatility differently). When McMillan asked what GJR was — he'd forgotten the acronym — the agent explained it immediately.

Step 5 — Output generation (07:07–08:55)

The agent didn't just print numbers. It produced a complete analytical package:

All from an upload McMillan hadn't seen before the call. All in under 10 minutes.

Live critique — the critic subagent

McMillan then spawned a second agent — a GLM 5.2 subagent — and gave it one instruction: review the GARCH agent's work and flag anything questionable.

The critic caught that these appeared to be actively managed equity returns and flagged a methodological concern: active management can introduce artificial serial dependence. Managers targeting stable volatility, using consistent rebalancing rhythms, or applying risk controls can suppress natural return clustering and distort autocorrelation patterns. This matters for GARCH, which assumes volatility dynamics are market-driven, not manager-driven. The GARCH agent hadn't raised this. The critic did.

Udayan watched an AI system catch a methodological issue that would matter in practice. This wasn't “statistical collation.” It was second-order reasoning.

From folder to dashboard

McMillan pointed out the agent's file system — empty at the start of the call, now populated with charts, markdown summaries, and Python scripts:

The GARCH agent's outputs are reusable code. Run it once, and you can run it on a cron job every month, or turn it into an interactive dashboard.

“This is on my computer?”

Then came the moment that changed the conversation. Udayan realized the demo was running on his own provisioned box.

Yes — his own provisioned box. By the end of the call, Udayan was logged into scottsdale.advisorclaw.ai himself, changing the appearance to light mode, increasing the font size, and exploring the file system and agents on his own.

What Udayan saw was not a vendor product demo. It was a live build on his own data. Here's what emerged.

GARCH volatility engine

The agent calibrated five GARCH variants on Udayan's Apex Equity Growth returns:

Autocorrelation & regime detection

Critic review layer

A second agent (GLM 5.2) reviewed the first agent's output and flagged:

The architecture

The transformation shows up in both the numbers and the dynamic of the meeting itself.

Why it landed

The pivot — from power user to fleet deployment

At lunch after the demo, Udayan shifted the conversation. His personal GARCH use case is real, and he was already thinking about how he'd use it. But the larger opportunity is getting AdvisorClaw in front of the RIA's 34 advisors.

Udayan was clear about what would work for his team. These advisors are relationship-driven professionals — they build trust with clients over years, not quarters. Their value is in understanding family dynamics, tax situations, and estate complexity, not in running volatility models. They need tools that amplify what they already do well, not tools that require them to become something else.

They operate in portfolio frameworks clients understand, translating market noise into plain language because that's what their clients need. Any new tool has to fit that workflow — not because the advisors can't handle complexity, but because their clients shouldn't have to.

The firm also positions itself as a multi-family office, and Udayan is always looking for tools that support that holistic approach. Many of his advisors came from wirehouses or smaller RIAs where they never had access to true family-office infrastructure. The right tools can help them deliver that experience without forcing them to become technologists.

Ben's response: don't lead with quantitative firepower. Build something any advisor can pick up immediately — simple on the surface, powerful underneath. Something that feels like an upgrade to their existing process, not a replacement of their expertise. Go through the advisor trainer, who knows what the team will actually use.

Udayan agreed. He's bringing the firm's advisor trainer to the next lunch to see what might work. The GARCH power user became the fleet-deployment champion.

This story isn't really about GARCH

Every sophisticated investor has a methodology. Udayan's is GARCH volatility modeling with walk-forward validation, Black-Litterman views matrices, and Markov-switching regime detection. Another CIO might lead with factor models, stress testing, or co-integration analysis. The point isn't which methodology — the point is that AdvisorClaw can build to any methodology.

Udayan spent months frustrated with generalist AI that couldn't remember, couldn't learn, and couldn't reason beyond internet consensus. He described what he wanted in his own words, in a 28-minute first call. He watched it get built live on his own data in under an hour. He logged into his own box during the demo.

The sequence of value delivery was compressed to a single meeting. No RFP. No procurement. No six-month pilot. A CIO articulated a sophisticated quantitative workflow, watched it materialize on his own strategy returns, and asked for the output by email — all within 54 minutes.

That's not just a feature request. It's the thesis of AdvisorClaw in the client's own language: a system that remembers your methodology, tracks your iterations, validates your models out of sample, catches its own errors, and gets smarter within your framework.

If a quantitatively skeptical CIO can see his ideal workflow built live on his own data in under an hour — what can you build on yours?