Multi-Source Financial Data Pipeline — SEC Filings, Investor Presentations, Transcripts (Python)

I need a pipeline built to collect and normalize recurring financial/operating metrics for ~40-50 public companies within the same industry vertical, pulled from four source types per company per quarter: SEC EDGAR XBRL filings (structured — data.sec.gov/api/xbrl/companyfacts/) Investor presentation decks (PDF — quarterly earnings slide decks) Press releases (earnings release exhibits, HTML/PDF) Earnings call transcripts (text, likely sourced via a transcript provider API or scraped from public sources) Why this is harder than a standard scrape: the same metric often needs a different source depending on the company and the quarter. Some metrics are XBRL-tagged (reliable, structured). Some only ever appear in the investor deck as chart labels or MD&A-style callouts (less reliable, prone to ambiguous formatting — e.g. stacked bar charts with unclear per-category values). Some are only stated verbally on earnings calls. I need a pipeline that: Tries the most reliable source first (XBRL) and falls back through deck → press release → transcript only when the metric isn't available upstream Tags every value with its source type, exact document, and confidence level — never silently blends a "reported exact" figure with an "extracted from a chart, uncertain" figure without distinguishing them Flags anything it can't resolve confidently rather than guessing (e.g., ambiguous chart segment values, contradictory figures across sources) — I'd rather see an explicit gap than a wrong number Scope of work: XBRL layer: Tag discovery + alias mapping across custom/standard tags (companies rename extension tags between filings — pipeline needs to handle this, not hardcode one tag name per metric). Document ingestion layer: Pull quarterly investor decks and press releases (PDF/HTML) per company; extract ~15-25 recurring metrics from tables and structured callouts. Flag anything sourced from ambiguous chart/graphic elements rather than a labeled table. Transcript layer: Given transcript text (source TBD — open to recommendations on providers), extract specific stated metrics/guidance figures via targeted NLP/LLM-based extraction, with the source quote retained for audit. Merge/precedence logic: Combine all four sources per metric per period with clear precedence rules and conflict flagging when sources disagree. Output: Normalized long-format table — company, metric, period, value, source_type, source_document, confidence_flag, source_excerpt (for audit). Out of scope for now: Front-end/dashboard, real-time alerting. This is a batch data pipeline. Ideal candidate has: Experience with SEC EDGAR XBRL API specifically (not just general scraping) PDF/table extraction experience (pdfplumber, camelot, or similar) — and comfort admitting when a chart/graphic can't be reliably parsed rather than guessing LLM-assisted extraction experience for unstructured text (transcripts, MD&A) with source-grounding/citation discipline Python, pandas, clean documented code To apply, please answer: Have you built pipelines combining structured (XBRL/API) and unstructured (PDF/transcript) financial data sources? Describe one. How do you handle a metric that's stated in a chart/graphic in a PDF where the exact value or category mapping is visually ambiguous from extracted text alone? What's your approach to flagging low-confidence extractions vs. silently returning a best guess? Rough time/cost estimate for a proof-of-concept covering 5 companies × 10 metrics across all four source types. Budget: Fixed price, milestone-based. Milestone 1: proof of concept on 5 companies/10 metrics across all four sources, before committing to full scope.