Portfolio Fit

Ask a question such as "how's my portfolio allocated?" or "is AMD returning well for me?" and the system builds the answer on the spot: charts and cards drawn from your holdings, narrated by a voice. None of it is pre-generated. Each question produces its own interface.

A classification model reads the sentiment and intent of the message and picks the visual that best explains it. Skeleton UI holds the place while the narration starts, so the response feels immediate.

Live interrupt

If a term loses you, such as "what is a point?", you can interrupt mid-explanation. The system explains the term, then lets you expand into another visual or resume the walkthrough exactly where you left off, with no restart.

Beyond the chat, a dedicated screen lets users choose an interest, like Technology, and see ranked picks. Each card carries a fit score against the user's current holdings, five-year average return, and plain-language reasons, such as adding international exposure to a U.S.-heavy portfolio with low overlap.

Where the fit score comes from

The score isn't a language model's opinion. JEV is asked a fixed list of typed questions about the user's portfolio, 19 per request, and for each one it returns a probability for every answer option we wrote. It never names a ticker and never writes free text. From there, three steps turn those answers into the cards you see.

1 · Classify

Score questions such as "how much new money should go into each asset class?" and choice questions such as "which sector should a new pick be in?" produce a plan. Code turns the answers into recommendation slots across assets like stocks, rental homes and bonds.

2 · Filter

Code narrows the pool of about 500 candidate assets to 20 using those answers, and JEV's confidence sets the order. The surest answer is applied first and strictly. An unsure one goes last, where the filter can widen it or skip it instead of forcing a weak guess.

3 · Score and rank

JEV rates each remaining candidate on a five-step expected-return scale, and code turns that into the fit score. Yield comes from its return-band probabilities. The best scorers fill each category's slots, and a second pass checks every reason on the card with a yes or no.

The split is deliberate: the model makes judgment calls, and code computes every number on screen. If the model service fails, each stage falls back to rules.

Slides from the team's pitch. Figures in the diagrams are illustrative.

Next.js and TypeScript on both the front and back end. GPT-4.1 was chosen for the voice path because response speed mattered most. ElevenLabs provides the narration. We couldn't train a model in 36 hours, so intent and sentiment classification runs on JEV, a zero-shot classifier that had just been released.

Stack & Tools

TypeScript Next.js & React GPT-4.1 ElevenLabs JEV Vercel