Inspiration
Financial risk assessment is notoriously slow, manual, and often biased by the narrow perspective of a single analyst. We were inspired to build a tool that automates the grunt work, reduces bias, and provides a more holistic, multi-perspective analysis to augment the capabilities of a junior analyst.
What it does
BerkshireBot (with the product dashboard "AgentSeer") is an autonomous "Analyst Swarm" for workplace financial analysis. When a user inputs a stock ticker (e.g., $TSLA), it spawns a team of specialist AI agents that work in parallel:
A "SEC Agent" reads regulatory filings.
A "News Agent" scans breaking news.
A "Social Agent" checks public sentiment.
A "Chart Agent" performs technical analysis.
An "Analyst Agent" reviews professional analyst ratings.
These agents conduct their research simultaneously. Their findings are then given to a "Governor Agent," which moderates a "debate" between their conflicting viewpoints. The final output is a single, balanced investment memo and risk assessment that fairly presents all sides, citing its sources.
How we built it
We built this on a multi-agent framework using LangGraph Studio as the central orchestrator.
Orchestrator (LangGraph): The main workflow begins when a user inputs a stock ticker. It uses LangGraph's state object to manage the analysis and "forks" the process into five parallel tracks for the specialist agents.
Specialist Agents: Each agent is a dedicated function:
SEC Agent: Connects to the SEC's EDGAR database, pulls the latest 10-K filing, and uses an LLM (Claude) to summarize the "Risk Factors" section.
News Agent: Scrapes news sites for recent articles and generates a high-level summary.
Social Agent: Scans Twitter and Reddit for social indications and favorability.
Chart Agent: Scrapes a stock chart image (e.g., from Yahoo Finance) and performs technical analysis and prediction on the visual data.
Analyst Agent: Scrapes and summarizes financial analyst reportings for the stock.
Governance and Synthesis:
Governor Agent: The graph "joins" at this node. This agent is prompted with Holistic AI principles ("be unbiased, cite sources, present conflicting views fairly"). It receives all specialist reports, moderates a final "debate" by prompting the LLM with all inputs, and writes a balanced investment memo.
Risk Assessment Agent: This final agent takes the Governor's memo and drafts a formal risk assessment.
The "Product" (AgentSeer): The final output is displayed on a dashboard that visually shows the entire pipeline—the "fork" to the individual agents, their key findings, and the final "joined" report, providing transparency into the model's reasoning.
Challenges we ran into
Synthesizing Conflicting Data: The hardest part was designing the Governor Agent. Prompting it to handle and fairly represent conflicting reports (e.g., a bullish "Chart Agent" vs. a bearish "SEC Agent") and moderate a "debate" to produce a single, coherent memo was a significant challenge.
Reliable Data Extraction: Building robust scraping agents that can consistently pull data from diverse and dynamic sources like the SEC EDGAR database, news sites, and social media feeds was difficult.
State Management: Effectively managing the state in LangGraph, especially when "forking" the analysis into parallel tracks and then "joining" all the disparate findings (text, summaries, image analysis) for the final governor node, required careful planning.
Accomplishments that we're proud of
🏆 Accomplishments that we're proud of We are proud of successfully implementing the full "Analyst Swarm" architecture using LangGraph, demonstrating a practical parallel "fork-join" workflow for complex analysis.
The Governor Agent is a key accomplishment. We successfully prompted it to integrate Holistic AI principles, forcing it to remain unbiased and present conflicting views fairly, which moves beyond a simple summary.
The "AgentSeer" dashboard provides crucial transparency. It allows a junior analyst to see how the final report was constructed, building trust by showing the individual findings that led to the conclusion.
What we learned
The "Swarm" is > the "Sum": We learned that a multi-agent "swarm" of specialists genuinely produces a more comprehensive and less biased analysis than a single, monolithic AI trying to do everything.
LangGraph is Perfect for this: LangGraph Studio is a powerful tool for building and visualizing complex, stateful agentic workflows, especially for parallel processing and iterative cycles like a "debate."
Governance is Essential: We learned that autonomous agents need a "governor" to be useful in a business context. The final debate, governed by principles of unbiased reporting, is the most critical step for producing a trustworthy, decision-ready report.
What's next for BerkshireBot
Add More Specialists: We plan to expand the swarm with new agents, such as a "Macro-Economic Agent" (to analyze interest rates, inflation, etc.) and a "Competitor Agent" (to analyze peer companies).
Human-in-the-Loop: We want to make the "debate" interactive, allowing the junior analyst to ask follow-up questions to the Governor or even challenge a specific specialist agent's findings.
Portfolio Analysis: We aim to scale the system to analyze an entire portfolio of stocks at once, not just a single ticker, and provide a holistic risk assessment for a user's entire holdings.
Log in or sign up for Devpost to join the conversation.