CoolFace
Apppublic

varunbunny/earnings-call-research-tool

sourceHugging Faceupdated 7mo agoView on Hugging Face
0likes
App README

Earnings Call Research Tool

This project implements a structured research tool for analyzing earnings call transcripts and management discussion sections. The system is designed as a fixed research tool (not a chatbot) and produces analyst-ready structured outputs.

It supports PDF, TXT and DOCX files and includes OCR fallback for scanned documents.

Supported Input

  • —Earnings call transcripts
  • —Management discussion / commentary sections
  • —File formats:
  • —PDF
  • —TXT
  • —DOCX

Scanned PDFs are supported using OCR.

Research Tool Implemented

Earnings Call / Management Commentary Summary

The tool produces the following structured outputs:

  • —Management tone (optimistic, cautious, neutral, pessimistic)
  • —Confidence level (high, medium, low)
  • —Key positives mentioned (with supporting quotes)
  • —Key concerns / challenges (with supporting quotes)
  • —Forward guidance:
  • —revenue outlook
  • —margin outlook
  • —capex outlook
  • —Capacity utilization trends
  • —New growth initiatives described

Each extracted point is backed by a direct quote from the transcript.

Hallucination Control

To avoid hallucinations:

  • —The model is restricted to use only the provided transcript.
  • —Every positive, concern and growth initiative must include a supporting quote.
  • —If a section is not present in the transcript, the tool returns null instead of generating content.

Handling Missing and Vague Information

  • —When guidance is vague, short descriptive phrases are returned (for example: “stable” or “moderate growth”), along with the supporting quote.
  • —If a data section is not mentioned in the transcript, it is returned as null.

System Overview

High-level pipeline:

  1. 1.Document upload
  2. 2.Text extraction
  3. 3.OCR fallback for scanned PDFs
  4. 4.Structured analysis using an LLM
  5. 5.Analyst-ready output
  6. 6.JSON and PDF export

Technology Stack

  • —Streamlit (UI)
  • —Python
  • —PyMuPDF, python-docx (text extraction)
  • —Tesseract + Poppler (OCR fallback)
  • —NVIDIA hosted LLM endpoint (OpenAI-compatible API)
  • —ReportLab (PDF export)

Deployment

The application is deployed as a Docker service and can be accessed via a public URL.

OCR dependencies (Tesseract and Poppler) are installed inside the container.

Limitations (Free Hosting)

  • —OCR on scanned PDFs may take longer to process.
  • —To remain within LLM context limits and improve reliability, only the first 8,000 characters of extracted text are analysed.
  • —Large files may result in higher latency.

Running Locally

bash
pip install -r requirements.txt
streamlit run app.py