Quant Research Agent

• 

  landing page

Inspiration

The Google Gemini 3 model card showed a major leap in reasoning, tool use, and agentic capabilities. At the same time we faced a painful problem where Quantitative research is slow and cognitively heavy when done manually. We thought what if we can utilize Gemini's capabilities to solve this painful process ?

What it does

QuantAgent turns a trading idea written in plain English into a complete research pipeline. It automatically generates a proposal, gathers and prepares data, engineers features, writes strategy code, runs backtests, evaluates performance, and outputs both explanations and a working trading system.

How we built it

We built QuantAgent as an agentic system powered by Gemini 3 with a standard quantitative research process with: Gemini 3 pro for deep reasoning Gemini 3 flash for fast structured tasks Function/tool calling to enforce strict backtest and evaluation outputs

Challenges we ran into

1. Limited libraries in Gemini’s Python sandbox (missing some finance tools)
2. Payload size limits when passing large market datasets
3. LLM hallucinations requiring strict schemas and validation guards

Accomplishments that we're proud of

1. Build a working end-to-end quantitative workflow
2. Creating a stable multi-step agent engine
3. Real backtest with executable strategy code
4. Build a strong auto optimization engine

What we learned

1. LLMs sometimes deviate on the desired goal , strong validation is required to achieve a stable agentic workflow
2. As LLMs are stateless , thought signatures helps the model on multistep reasoning building a knowledge graph
3. Guiding the model by tool calls reduces hallucination drastically

What's next for Quant Research Agent

1. Custom python sandbox for supporting complex and custom financial tools
2. Real time optimization and adapting to changing market regimes
3. Larger dataset support

Built With

• alpha-vantage
• gemini-3
• react