Case Study

Portfolio AI

Building AI that knows what it knows—and what it shouldn't decide.

A hands-on exploration of how deterministic systems, AI reasoning, and human judgment can work together to produce trustworthy decision intelligence.

AI Product StrategyHands-on BuildLLM EvaluationDecision Intelligence

3

Experiments

78

Offline tests

2

Controlled A/B runs

Evidence-grounded
reasoning

The Product Thesis

Establish Truth

Deterministic systems
establish what is true.

Create Understanding

AI reasons about
why it may matter.

Human Judgment

Humans decide.

Establish Truth

Deterministic systems
establish what is true.

↓

Create Understanding

AI reasons about
why it may matter.

↓

Human Judgment

Humans decide.

Portfolio AI started as a portfolio-analysis idea, but evolved into an exploration of how to divide responsibility between software, AI, and humans—so AI can produce useful decision intelligence without inventing financial facts, investor policy, or decisions.

The Architecture

A deliberate division of responsibility.

I separated calculation from reasoning so each component owns the responsibility it is best suited to handle.

Establish Truth

Portfolio Data

↓

Deterministic Analytics

↓

Exposure Intelligence

+

Investor Intent

↓

Deterministic Comparisons

Create Understanding

AI Reasoning

Synthesize trusted facts

Prioritize insights

Explain implications

Surface uncertainty

Evidence validation

Human Judgment

Intent

Preferences

Trade-offs

Policy

Decisions

Trust comes partly from deciding which component should own which responsibility.

Experiment 01

A good AI answer
that I rejected.

The first live AI analysis was useful. It connected portfolio facts and produced meaningful reasoning.

But the system rejected the response because its evidence references could not be reliably validated against the trusted portfolio context.

USEFUL AI ANSWER

Meaningful portfolio insights

TRUST CHECK FAILED

Evidence references violated the required contract

PRODUCT DECISION

Fix the AI contract—not the validator

NEXT RUN

Validation passed

Technical detail

incorrect: /snapshot/top_holdings/0/portfolio_weight
correct: /top_holdings/0/portfolio_weight

“A plausible AI answer isn't the same thing as a trustworthy AI product.”

Experiment 02

I made the AI smarter
without changing the model.

Human evaluation exposed a deeper problem: the model knew which securities were held, but not necessarily which companies the investor was actually exposed to.

Instead of asking the LLM to infer ETF composition, I built deterministic ETF look-through and gave the reasoning layer better trusted facts.

NVIDIA · Synthetic Evaluation Portfolio

Instrument-level view

6.1051%

Direct exposure

Exposure Intelligence

7.7448%

Known effective exposure

Direct · 6.1051%
ETF-mediated · 1.6397%

$33,300 direct + $8,944 ETF-mediated = $42,244 known exposure

What improved

The same reasoning layer could now connect direct and ETF-mediated company exposure.

What remained unknown

Only 14.1045% of ETF value was attributed in the deliberately partial synthetic dataset.

Known exposure ≠ complete exposure

“Better reasoning sometimes requires better facts—not a better prompt.”

Experiment 03

The AI knew what I owned.
It didn't know what mattered.

Exposure Intelligence answered a better question: what am I actually exposed to? But knowing that NVIDIA represented 7.74% of known portfolio exposure still didn't tell the system whether that number deserved attention.

Instead of letting the model invent investment policy, I modeled investor-defined intent explicitly and compared portfolio facts against it deterministically.

Synthetic Investor Context

Maximum company weight

7.0%

Reported crypto allocation range

0–10%

Arbitrary synthetic evaluation inputs—not investment recommendations or suitability standards.

NVIDIA

Known exposure: 7.74%

Investor-defined guideline: 7.0%

Status: KNOWN EXCEEDS GUIDELINE

KNOWN EXCEEDS GUIDELINE

Even with partial ETF coverage, known exposure already exceeds the stated synthetic guideline. Additional exposure cannot undo the known exceedance.

Alphabet

Known exposure: 6.42%

Investor-defined guideline: 7.0%

Status: INDETERMINATE

INDETERMINATE

Known exposure is below the guideline, but incomplete ETF coverage means the system cannot conclude that total modeled exposure is within it.

Same guideline. Different evidence.
Different conclusion.

The deterministic comparison preserves the difference between what is known and what remains uncertain. The AI reasons over that state rather than manufacturing certainty.

What the Experiments Taught Me

Trustworthy AI required improving more than the model.

Across the three experiments, the biggest improvements did not come from changing the model. They came from improving the system around it—first the contract, then the facts available to the AI, and finally the context needed to understand what matters to the investor.

Better Contract

Make AI reasoning verifiable.

The first experiment showed that useful output is not enough. AI reasoning needs a clear evidence contract so the product can verify what the model is claiming.

Trust requires traceability.

↓

Better Facts

Give AI better things to reason over.

ETF look-through moved company exposure into deterministic analytics instead of asking the model to infer it. The same reasoning layer became more useful because its trusted context became better.

Better reasoning can start with better facts.

↓

Better Context

Model what matters to the human.

Investor intent turned portfolio facts into decision context. The system could compare what was known with investor-defined guidelines without asking the AI to invent policy.

Personalization should not require guessing.

The progression became clear:
Better contract → Better facts → Better human context.

The role of the AI stayed deliberately constrained: synthesize trusted information, explain implications, surface uncertainty, and help the human investigate. Calculation, policy, and the final decision remained outside the model.

Where Portfolio AI Goes Next

From trusted reasoning to richer decision intelligence.

The experiments established a foundation: portfolio facts are calculated deterministically, investor intent is modeled explicitly, and AI reasons over trusted context rather than inventing it.

The next step is to expand that trusted context so the system can answer progressively richer questions about a portfolio.

01 — Account Intelligence

Where do I own it?

Understand how exposure is distributed across accounts and surface account-level context relevant to future decisions.

02 — Market & Regime Intelligence

What is changing around my portfolio?

Connect portfolio exposure with trusted market signals and changing conditions while preserving uncertainty.

03 — Scenario Intelligence

What if something changes?

Use deterministic scenario engines to calculate hypothetical outcomes, then let AI explain the implications.

04 — Opportunity Intelligence

What opportunities exist within what I already own?

Surface opportunities such as income, diversification, or allocation considerations from trusted analytics while keeping the final decision with the investor.

The architecture stays the same even as the questions become more sophisticated.

Trusted data→Deterministic intelligence→AI reasoning→Human decision

Closing note

Building at the intersection of AI reasoning, deterministic systems, and human decision-making.

GitHub ↗LinkedIn ↗Back to home →