Équipe 35

Inspiration

We built AlphaForge around a question an investor should be able to answer when they see a portfolio: Why did I get these weights? We wanted people to trace an allocation back to the goals, cash needs, and restrictions they shared, then see how those inputs affected the result.

What it does

AlphaForge starts with a guided investor interview. It asks about goals, time horizon, comfort with risk, ability to absorb losses, and investment restrictions. When answers conflict, the interview asks the investor to clarify before confirming their profile.

After confirmation, Portfolio output shows the limits the optimizer can apply. For example, a $1,000 reserve on a $10,000 portfolio becomes a 10% minimum cash weight. The report shows holdings, estimated returns, risk metrics, historical growth versus SPY, and drawdown. Its Reasoning and Math tabs explain which limits were binding and show the formulas with the calculated values.

How we built it

We built the interview and dashboard in Next.js and TypeScript. The app sends structured calculation requests through an AWS Bedrock AgentCore integration. The Python optimizer uses CVXPY to maximize estimated monthly return after trading costs, subject to the confirmed limits.

$$ \max_{w,c} \quad \mu^\top w + r_c c - \kappa \lVert w - w_{\mathrm{prev}} \rVert_1 $$

Here, $w$ represents asset weights, $c$ is the cash weight, and the final term estimates the cost of changing the asset weights. We used Recharts for the plots and KaTeX to render the equations in the report.

Challenges and what we learned

The hardest part was translating interview answers into rules the optimizer could justify. An explicit crypto cap maps to a numeric limit. “I’m cautious” calls for a follow-up before it can define a risk limit. Fund-level exclusions also need holdings data to verify. We learned to show which rules the solver can enforce and where the available evidence falls short.

We also found a deployment mismatch during testing. The deployed AWS runtime rejected the newer calculation fields. We verified the full flow locally with the current Python backend, from completing the interview to reopening the saved report with charts and worked formulas.

Built With

Share this project:

Updates

Submission history