EquityPulse

"The Hedge Fund in Your Pocket." Autonomous Multimodal Multi-Agent Platform (Gemini 3.0 Pro). A swarm of agents debates P&L,BS, Macro, & Vision data to generate an institutional investment thesis.

Comment

Project Name

EquityPulse: The Autonomous AI Investment Committee

Tagline (Short Description)

"The Hedge Fund in Your Pocket." Autonomous Multimodal Multi-Agent Platform (Gemini 3.0 Pro). A swarm of 6 agents (incl. Management) debates 10-Ks, Macro, & Vision data to generate institutional investment thesis.

Inspiration

I have been investing in the equity markets for more than 8 years. Despite reading every book and watching every video, I found myself unable to beat a simple index fund. I realized it wasn't just me: 91% of retail investors lose money.

Why? Because we are fighting a war we cannot win.

It’s not about P/E ratios or simple math. The real money in the stock market is made on Qualitative Intuition:

  • Detecting looking at a CEO's hesitation during an earnings call.
  • Understanding a "Sector Rotation" from Tech to Industrials before it shows up on the charts.
  • Sensing the "Macro Vibe" of a geopolitical shift.

These are NOT quantitative terms. They are deeply qualitative, requiring decades of experience and teams of 50+ analysts working 100-hour weeks. As a retail investor with a day job, it is physically impossible for me to process this ocean of unstructured data.

I realized I wasn't losing because I couldn't do the math. I was losing because I was deaf to the Human and Macro story behind the ticker.

The "Deep Learning" Dead End Years ago, I tried building this using LSTMs and Core Machine Learning algorithms (not Transformers). It didn't work. Why? Because investing is a game of probability, not just points. Qualitative nuances like "Management Tone" cannot be captured by simple regression. We needed a Reasoning Engine, not a pattern matcher. Gemini 3.0 Pro was the breakthrough that finally allowed us to ingest entire 10-Ks.

The "10,000 Hour" Problem Mastering the old way isn't just hard; it's impossible for most of us. It takes decades of "Fail, Learn, Fail, Learn" to truly understand sector cycles and management forensics. We don't have decades or the tuition fee of a dozen blown accounts. We need institutional-grade research today.

That's why I built EquityPulse. It bridges the gap between "Retail Guesswork" and "Institutional Wisdom."

How we built it

EquityPulse is not a chatbot. It is an Autonomous Investment Committee. We orchestrated a swarm of 6 specialized agents using LangGraph and Gemini 3.0 Pro:

  1. Fundamental Agent (The Buffett): Reads the entire 10-K to find Moats.
  2. Quant Agent (The Math Whiz): Detects fraud and bankruptcy risk (see Math section).
  3. Risk Agent (The Bear): Hunts for geopolitical headwinds.
  4. Technical Agent (The Trader): Times the entry.
  5. Sector Agent (The Macro Strategist): Analyzes tailwinds/headwinds.
  6. Management Agent (The CIO): Resolves conflicts using a "Conflict Engine."

We used Gemini 3.0 Flash for a sub-100ms voice loop, allowing users to talk to their committee while driving.

Challenges we ran into

1. The Trust & Consistency Crisis Early versions gave different answers every time. That’s a dealbreaker. We solved this by:

  • Chain-of-Thought (CoT): Forcing agents to "show their work" step-by-step.
  • Validation Agents: Critics that strictly cross-check every number against the API.
  • More Vectors: Increasing inputs to 50+ parameters to triangulate the truth. This moved us from "vibes" to Deterministic Consistency.

2. Implementing Institutional Math To make the "Quant" agent credible, we needed real forensic accounting. Implementing the Beneish M-Score required complex multi-step math:

$$ M = -4.84 + 0.92(DSRI) + 0.528(GMI) + 0.404(AQI) + 0.892(SGI) + 0.115(DEPI) - 0.172(SGAI) + 4.679(TATA) - 0.327(LVGI) $$

Getting the LLM to reliably extract these 8 variables without hallucinating was a massive engineering challenge.

3. The "Vague Answer" Trap Simple bots say "It depends." We built a "Conflict Engine" where the Management Agent forces the Bull (Technical) and Bear (Fundamental) to debate, rendering a final, nuanced verdict.

Accomplishments that we're proud of

  • "Full-Spectrum Reasoning": We moved beyond simple summarization to actual Forensic Analysis of qualitative signals.
  • The UI: A "Dark Data" terminal that visualizes the Agent Traces in real-time, showing the user exactly how the decision is being made.
  • Multimodal Logic: Uploading a screenshot of a news article and having the agent analyze its specific impact on the portfolio.

What we learned

  • Specialization is everything: A generic "Finance Bot" fails. A specific "Short-Seller Persona" finds risks that others miss.
  • Qualitative > Quantitative: The math is easy. The story—management integrity, sector cycles, macro shifts—is where the alpha is. Gemini 3.0 Pro's 1M context window allowed us to finally capture that story.

What's next for EquityPulse

  • Real-time Portfolio Connection: analyzing your actual holdings.
  • Global Markets: Expanding to NSE (India) and European markets.
  • Proactive Alerts: Notification when your stock's "Z-Score" drops below safe levels.

Built With

  • cot
  • fastapi
  • financial-apis
  • gemini-3
  • gemini-3.0-flash
  • gemini-3.0-pro
  • langfuse
  • langgraph
  • multi-agent-orchestration
  • multi-model
  • postgresql
  • python
  • react
  • speach-to-text
  • stt
  • supabase
  • vision-model
  • yfinance
Share this project:

Updates