stock-research-pipeline

Stock Research Pipeline

Automate end-to-end equity research: screener.in data collection, NotebookLM-powered analysis with dynamically generated queries, and a professional PDF report with variant perception scorecard.

Prerequisites

• nlm CLI installed and authenticated (nlm login --check should succeed)
• Python venv at .venv/ with reportlab installed
• Internet access for screener.in and BSE PDF downloads

If nlm login --check fails, tell the user to run nlm login first and stop.

Input

The user provides a stock SYMBOL (e.g., GRAVITA, TCS, RELIANCE). Optionally they can specify:

• Number of quarters (default: 4)
• Specific quarters to analyze

Workflow

Step 1: Verify NLM Authentication

nlm login --check

If this fails, stop and ask the user to authenticate first.

Step 2: Fetch Concall & Presentation Links from Screener

Use WebFetch on the screener.in consolidated page (fall back to standalone if needed):

URL: https://www.screener.in/company/{SYMBOL}/consolidated/
Prompt: Extract ALL conference call transcript links AND investor presentation links.
        For each, return the date (month and year), type (Concall or Investor Presentation),
        and the full URL. Only include bseindia.com PDF links.
        Return as a markdown table: Date, Type, Link

If consolidated returns no results, retry with:

https://www.screener.in/company/{SYMBOL}/

Sort results by date (newest first). Select the most recent N quarters (default 4).

Step 3: Auto-Download PDFs

Create directories and download without asking:

mkdir -p data/companies/{SYMBOL}/concalls data/companies/{SYMBOL}/presentations

Download concalls:

curl -sL -o "data/companies/{SYMBOL}/concalls/{SYMBOL}_concall_{YYYY-MM}.pdf" "{URL}"

Download investor presentations:

curl -sL -o "data/companies/{SYMBOL}/presentations/{SYMBOL}_investor_pres_{YYYY-MM}.pdf" "{URL}"

Verify downloads with ls -lh. If any file is 0 bytes or suspiciously small (<10KB), warn and retry.

Step 4: Create NotebookLM Notebook & Upload Sources

nlm notebook create "{COMPANY_NAME} - Equity Analysis"

Save the notebook ID. Then upload each PDF:

nlm source add {NOTEBOOK_ID} --file "path/to/file.pdf"

Upload ALL downloaded files (both concalls and presentations). Wait for each to complete before starting the next.

Step 5: Fetch Screener Financial Data & Upload to NLM

This step extracts comprehensive financial data from screener.in, converts it to a PDF, and uploads it as an additional NLM source so that all queries can reference actual financial numbers.

Step 5a: Extract screener financial data

Use WebFetch on the screener page to extract ALL financial data:

URL: https://www.screener.in/company/{SYMBOL}/consolidated/
Prompt: Extract ALL financial data from this page in a structured format. Include:
        1. Company overview and description
        2. All key ratios (PE, PB, ROE, ROCE, Dividend Yield, etc.)
        3. Quarterly results table (all quarters visible) - Revenue, Expenses, Operating Profit, OPM%, Net Profit, EPS
        4. Profit & Loss annual data (all years visible)
        5. Balance Sheet data (all years visible)
        6. Cash Flow data (all years visible)
        7. Key financial ratios table (all years visible)
        8. Shareholding pattern (promoter, FII, DII, public)
        9. Peer comparison table if available
        10. Pros and Cons listed on the page
        Return everything as clean markdown with tables preserved.

If consolidated returns 404, fall back to standalone URL.

Step 5b: Save as markdown

Save the extracted data to data/companies/{SYMBOL}/screener_data.md for reference.

Step 5c: Convert to PDF and upload to NLM

NLM only accepts PDF sources, so convert the screener data to PDF using reportlab:

• Create a clean PDF with all financial tables (quarterly results, annual P&L, balance sheet,

cash flow, ratios, shareholding pattern)

• Include key metrics, pros/cons, and growth rates
• Save to data/companies/{SYMBOL}/{SYMBOL}_screener_data.pdf
• Upload to the NLM notebook:

nlm source add {NOTEBOOK_ID} --file "data/companies/{SYMBOL}/{SYMBOL}_screener_data.pdf"

This ensures NLM queries can cross-reference concall commentary with actual reported numbers, enabling much richer analysis (e.g., validating management guidance against actual delivery, tracking margin trends, balance sheet changes, shareholding shifts).

Step 6: Build Company Context (Critical Step)

This step is what makes queries dynamic and tailored. Gather context from TWO sources:

Source A: Screener page context (already extracted in Step 5a — reuse it, don't re-fetch)

Source B: NLM notebook describe

nlm notebook describe {NOTEBOOK_ID}

This returns an AI-generated summary of the uploaded sources, including suggested topics.

Step 7: Generate Tailored Analysis Queries

Using the company context from Step 6, generate 6 analysis queries. These are NOT hardcoded templates - they must be tailored to the specific company's business model, sector, and what's actually discussed in the sources.

The 6 query categories are:

1. Business Model & Value Chain - How does THIS specific company make money? What are the

unit economics? Adapt to the actual business (e.g., processing spread for a recycler, subscription metrics for SaaS, same-store growth for retail, NIM for a bank).

1. Industry Structure & Competitive Positioning - What's the industry structure for THIS

sector? Who are the real competitors? What moats exist? Adapt terminology to the sector (e.g., "organized vs unorganized" for Indian manufacturing, "TAM penetration" for tech, "branch network" for banking).

1. Management Quality & Execution - Track guidance vs delivery across the quarters available.

How has tone changed? What strategic bets are being made? Are they delivering on promises?

1. Financial Deep Dive - Quarter-by-quarter P&L, margins, balance sheet, return ratios.

Focus on the metrics that matter most for THIS type of business (e.g., EBITDA/ton for commodity processors, ARPU for telecom, book value for banks, SSG for retail).

1. Growth Triggers & Variant Perception - SOIC-style analysis. Scan for all standard growth

drivers (new products, geographic expansion, market share gains, capex, regulatory tailwinds, etc.). Identify where market consensus may be wrong. Build a VP scorecard with kill-switches.

1. Risks & Bull/Base/Bear Scenarios - What could go wrong? Build three FY+2/3 scenarios

using management's own guidance as the base case, then stress-test up and down with specific numbers.

Query generation guidelines:

• Each query should be 150-300 words, structured with numbered sub-questions
• Always ask for specific data points, numbers, and direct management quotes
• Reference sector-specific metrics and terminology
• Ask for quarter-by-quarter tracking where relevant
• Reference screener financial data explicitly (e.g., "Using the screener financial data, trace

the revenue journey from X to Y" or "Cross-reference the quarterly P&L from screener with management commentary") — this ensures NLM uses the uploaded screener source

• End each query requesting structured data presentation

Step 8: Query NotebookLM

Run each query sequentially using nlm notebook query:

nlm notebook query "{NOTEBOOK_ID}" "{QUERY}" -t 180

For the first query, capture the conversation_id from the JSON response. Use it for subsequent queries with -c {CONVERSATION_ID} to maintain context continuity.

Important: Parse the JSON output from each query and extract the answer field. Store all 6 answers for the report generation step.

If a query fails or times out, retry once. If it fails again, note the gap and continue with remaining queries.

Step 9: Generate PDF Report

Use