Generative‑AI assistant that automates core equity research tasks — from sourcing disclosures and news to synthesizing insights and drafting analyst‑ready output.
- Accelerate time‑to‑insight: Target ~75% reduction in initial research and report drafting time.
- Scale analyst coverage: Enable up to 3× more companies per analyst through workflow automation.
- Improve decision consistency: Standardized analysis and templated outputs reduce variance and rework.
- Ingest: Retrieve company disclosures (10‑K/Q), earnings transcripts, news, and pricing/ratio context.
- Summarize: Distill key sections (business overview, risks, segments) with reliable citations.
- Sentiment: Score narrative signals across filings and news to detect trend/inflection.
- Valuation aides: Surface ratio snapshots (P/E, P/S, EV/EBITDA) and DCF inputs for review.
- Draft: Generate an editable investment brief (talking points, risks, watch‑items) for analysts.
Example prompt:
Analyze AAPL and prepare a 1‑pager with overview, recent drivers, valuation context, and top 5 risks, citing important passages.
- Orchestration / UI: CLI + optional Streamlit/Flask front‑end
- GenAI runtime: Azure AI Foundry (model deployments, safety, monitoring)
- Retrieval: Vector store (Azure AI Search or pgvector) for Retrieval‑Augmented Generation (RAG)
- Pipelines: ETL jobs to fetch and chunk source docs; metadata & citation tracking
- Guardrails: Content filters, source‑attribution, confidence flags, human‑in‑the‑loop review
- Tests: Unit tests for agents, E2E smoke tests on a small ticker set
app/
├─ ui/ # Streamlit / Flask app (optional)
├─ agents/ # retrieval, synthesis, drafting
├─ pipelines/ # ingestion, chunking, indexing
├─ services/ # Azure, vector DB, storage
├─ eval/ # prompt evals, regression tests
└─ tests/ # unit + E2E
configs/
├─ app.toml # feature flags
└─ connections.toml # endpoints, collections, index names
data/
├─ raw/ # downloaded source docs (gitignored)
└─ index/ # vector artifacts (gitignored)
- Azure AI Foundry (model deployments & safety)
- Python (pandas, requests, pydantic, fastapi/streamlit)
- Vector DB: Azure AI Search or PostgreSQL + pgvector
- Scheduling: cron / GitHub Actions for ingestion refresh
- Testing: pytest + data‑driven prompt checks
- Cited answers: Every claim links back to a paragraph in source docs.
- Configurable templates: 1‑pager, 3‑pager, or deck outline.
- Valuation helpers: Ratio snapshots and scaffolds for DCF inputs (manual review encouraged).
- Sentiment tracks: News vs. filings sentiment deltas to spot narrative shifts.
- Analyst controls: Redline edits, risk tagging, and watch‑list export (CSV/Markdown).
- Company overview: business model, segments, geography
- Recent drivers: product launches, guidance shifts, regulatory updates
- Valuation context: P/E, P/S, EV/EBITDA time‑series snapshot
- Top risks: sourced from Item 1A and recent transcripts
- Retrieval‑only answers for factual claims; no free‑form “knowledge” without a source
- Guardrails for sensitive content; explicit confidence flags on low‑evidence sections
- Human‑in‑the‑loop: Analysts approve draft output before publication
- Test set of tickers ensures prompt/output regressions are caught early
- Broker transcript and Q&A slot extraction
- KPI extraction by sector (e.g., DAUs/MAUs for internet, RPO for SaaS)
- Valuation table auto‑refresh via scheduled data pulls
- Redteam evaluations and bias checks
- Multi‑company comp‑table generation
Built as part of an MSBA project exploring how GenAI can automate and standardize equity research while keeping analysts in control of judgment and sign‑off.