Pharma Case

Inspiration

Pharmaceutical companies release enormous amounts of financial information through SEC filings, earnings reports, and other disclosures. Hidden within those documents can be important details that change how investors understand a company's performance.

We built Pharmer Research to make financial investigations faster, more accessible, and grounded in verifiable evidence. Our goal was to use specialized AI agents to uncover meaningful findings and turn them into actionable recommendations for pharmaceutical executives.

What it does

Pharmer Research is a multi-agent AI research platform that investigates pharmaceutical companies.

Users select a company, and the system:

Collects annual and quarterly financial filings.
Uses specialized AI research agents to investigate financial disclosures.
Identifies and verifies evidence-backed findings with citations to original documents.
Quantifies potential financial impacts and ranks eligible findings.
Selects a key pharmaceutical product and compares available evidence about its competitors.
Develops a strategic recommendation for the CEO, including a 24-month implementation plan.
Generates a downloadable PDF research report.

Our goal is to turn complex financial disclosures into understandable, evidence-backed strategic insights.

How we built it

We developed Pharmer Research primarily in Python, using the Anthropic Claude API for specialized AI research and reasoning.

The system uses SEC EDGAR to retrieve company filings and XBRL data to extract structured financial information. Our Librarian prepares the evidence corpus, Hunter agents investigate it, the Fact Checker verifies citations and financial claims, and Quant calculates financial impacts using reproducible Python arithmetic.

Additional components handle deduplication, scoring, product comparison, strategic synthesis, CEO recommendations, and PDF report generation.

We built the interface using Streamlit and organized development around shared data contracts so multiple agents could work together without losing evidence traceability.

Challenges we ran into

One of our biggest challenges was making sure AI-generated findings were actually supported by financial evidence.

Different filings presented numbers with different rounding precision, table structures, and reporting periods. We built verification logic to reconcile those differences without treating unsupported numbers as correct.

We also encountered API authentication and workspace configuration problems, inconsistent financial assumptions across model runs, and integration challenges between independently developed pipeline components.

These issues pushed us to improve testing, source attribution, error handling, and financial reproducibility.

Accomplishments that we're proud of

We successfully built and demonstrated a working multi-agent research pipeline capable of investigating real pharmaceutical filings and producing a structured PDF report.

Our system includes citation verification, independent financial number-tie checks, quantitative impact calculations, and a process for turning research findings into strategic recommendations.

One particularly important achievement was proving that changes in verified financial-impact estimates could influence the ranking of findings, rather than relying on arbitrary ordering.

What we learned

We learned that building an effective AI research system requires more than connecting several language models.

Evidence provenance, accurate calculations, careful agent coordination, and transparent uncertainty are essential. We also learned why combining AI reasoning with deterministic verification is especially important for financial analysis.

What's next for Pharmer Research

We plan to expand pharmaceutical evidence coverage, strengthen product and competitor comparisons, and improve our evaluation of whether findings have already been reflected in market expectations.

We also want to improve the stability of financial assumptions, expand international-company coverage, and make the platform easier for nontechnical users to explore research findings.