Inspiration

We wanted to explore a question: what happens when multiple AI agents work together to research trading strategies and manage a portfolio?

That became HellaBandz, a stock and ETF research platform that makes the trading process visible—from an agent’s initial hypothesis to its decisions, execution, and portfolio impact.

What it does

HellaBandz turns quantitative research into an interactive experience. Users can replay historical market data, watch agents evaluate opportunities, and inspect the reasoning behind their decisions.

Agents submit trading intentions to a shared coordinator. It reconciles competing requests, checks available capital and risk limits, and routes aggregate orders to an internal paper central limit order book (CLOB).

The dashboard displays strategy activity, executed trades, portfolio performance, and each agent’s contribution. Users can download a detailed simulation history for further analysis.

How we built it

We built the frontend with Next.js, React, and TypeScript, and the backend with Python, FastAPI, PostgreSQL, Solana, and TimescaleDB. The application runs through Docker.

We integrated Google Gemini for strategy proposals and commentary, with automatic OpenAI fallback to keep runs moving if the primary provider fails. Deterministic strategy rules and portfolio controls govern execution.

Our research loop develops and evaluates strategy candidates across separate data periods. Promotion checks determine which strategies receive capital, while a passive core provides a default allocation.

Historical stock and ETF prices come from Alpaca SIP snapshots. Our internal paper order book models liquidity from historical price bars so we can study coordinated execution without risking real money.

Challenges we faced

The biggest challenge was getting multiple agents to operate as one coherent trading system. Agents can disagree about the same stock, compete for available capital, or generate unnecessary trading activity.

We addressed this with shared allocation controls, intent netting, and checks before order submission. This gives the swarm a single execution authority while preserving each strategy’s individual contribution.

Another challenge was keeping accounting consistent across asynchronous services. We used a hash-chained event ledger and replay verification to check that recorded trades reproduce the stored portfolio state.

We also designed the research and execution components around limited compute resources, balancing strategy evaluation, execution, and model calls within a small VM budget.

What we learned

We learned how to connect generative AI with deterministic execution: models help formulate ideas and explain observations, while explicit rules govern orders, capital, and risk.

We also learned that coordination is a core part of multi-agent design. Giving agents a shared market view, clear budgets, and one execution pathway makes their behavior easier to understand and evaluate.

Working with Solana taught us how on-chain commitments can support verification of an off-chain application’s records.

Building HellaBandz strengthened our experience with quantitative research, event-driven systems, model failover, and interactive financial dashboards.

Accomplishments we’re proud of

Built an end-to-end multi-agent research and paper trading platform.
Integrated Gemini with automatic OpenAI failover.
Coordinated competing strategy intentions through a shared execution engine.
Added a Solana Devnet audit layer for verifiable record commitments.
Made agent activity inspectable through detailed simulation exports.
Validated the system through automated tests, browser checks, and a full historical replay with zero ledger rejections or reconciliation breaks.

What’s next

We plan to expand the strategy library, evaluate more market conditions, and improve execution realism. We also want to make the research loop faster and more adaptive while keeping it practical to run on a small virtual machine.