README.md

metadata

title: Amplegest
emoji: 📊
colorFrom: green
colorTo: gray
sdk: streamlit
sdk_version: 1.49.1
app_file: app.py
pinned: false

Amplegest — Post-Earnings Decision Intelligence

Amplegest turns public-company disclosures into a compact, evidence-linked decision memo for portfolio managers and senior equity analysts.

The product is designed for the first review after an earnings release: what changed, what confirms or challenges the thesis, which variable can swing the next estimate revision, and what must be monitored next. It does not issue a buy/sell recommendation or a price target.

Decision stack

Surface	Role in the workflow
Company Overview	Concise company onboarding: business model, revenue engines, geographic exposure, three-year evolution, investor-attention themes, relative price events, recent news, and the variables to watch.
PM Flash	One-screen decision memo: filing date and source coverage, experimental AI read-through, thesis-confirming and thesis-challenging evidence, the swing factor, and the next items to watch.
Evidence & Deltas	Audit layer: source-backed claims, before/after evidence, heuristic period deltas, risks, management commentary, and experimental AI hypotheses.
Financials	Historical KPI trends, guidance history, earnings history, and exportable structured data.
Ask AI	Cross-cutting utility for grounded follow-up questions. It supports the review; it is not the investment conclusion.

Evidence contract

Amplegest deliberately separates three different output classes:

Output class	Meaning	UI treatment
Source-backed fact	A factual claim tied to a 10-K, 10-Q, transcript, or news excerpt.	Source, period, item-level reliability, and evidence remain inspectable.
Heuristic delta	A deterministic before/after computation produced by a rule-based detector, such as a new risk phrase, dropped KPI, or language shift. The inputs are evidence; the detector's significance is a heuristic.	Labelled Heuristic · validate and shown with the underlying comparison.
Experimental AI hypothesis	A synthesis or interpretation, including the PM read-through, analytical tensions, between-the-lines reads, and earnings-quality assessments.	Labelled AI · experimental and presented as a question to validate against cited evidence.

Item-level reliability metadata describes source provenance and corroboration. The UI renders it as plain-language confidence, not as a HIGH / MEDIUM / LOW portfolio score, and never promotes it into a conviction indicator.

Fail-closed presentation policy

Amplegest does not display a market-derived output merely because a payload contains a value. The presentation layer hides the following by default:

• aggregate sentiment;
• any aggregate HIGH / MEDIUM / LOW reliability roll-up or percentage;
• D1 / D5 post-earnings returns;
• actual-versus-consensus comparisons.

These outputs remain hidden until their required alignment is explicit and testable: the same fiscal period and metric definition for actuals and consensus, the exact release timestamp and trading session for event returns, and a documented calibration methodology for aggregate scores. Missing or uncertain alignment means do not render.

This policy does not remove item-level source tags, evidence excerpts, or reliability labels from source-backed claims. Legacy briefs without a deterministic evidence_coverage result are shown as Legacy / unverified; the presence of a 10-K, 10-Q, or transcript tag alone never upgrades them to verified.

Architecture

• Offline ingestion — SEC EDGAR XBRL into SQLite; 10-K/10-Q Business, Segments/Geography, MD&A, Risk Factors, and earnings-call transcripts into the evidence stores.
• Company Overview — deterministic filing/call retrieval is injected before the agent loop; the source-fingerprinted English overview is emitted and verified within the brief's single synthesis call, while prices and public news refresh independently at display time.
• Runtime agent — a LangGraph tool loop retrieves structured financials, filings, transcripts, public news, and analyst data before one synthesis call produces the validated Pydantic brief and Company Overview.
• Retrieval — vector search over-fetches candidates and reranks them with a cross-encoder.
• Post-synthesis controls — deterministic source-reliability rules, corroboration checks, and heuristic period-delta detectors run after model synthesis.
• Decision UI — Company Overview onboards; PM Flash frames the decision; Evidence & Deltas provides auditability; Financials provides depth; Ask AI is the final cross-cutting tool.

See WRITEUP.md for product and methodology detail and ARCHITECTURE.md for data lineage and presentation controls.

Run locally

pip install -r requirements.txt
cp .env.example .env
python ingest.py AAPL --full
streamlit run app.py

Set the required API keys in .env first. A full re-ingestion is required for databases created before the evidence-lineage upgrade: legacy vector chunks and metrics without verified period context are intentionally hidden rather than silently upgraded.

For an already lineage-compatible database, python ingest.py AAPL is enough to backfill the Company Overview’s 10-K Business and Segments/Geography sections; the delta path reuses any transcript already stored.

Deploy to the Hugging Face Space

The Space serves the committed data/ snapshot through Git LFS. After any local ingestion, commit data/, then run python deploy_check.py; only run git push hf main:main when the preflight passes. Space runtime storage is ephemeral, so runtime-written caches and saved briefs do not survive a restart.

Intended use

Amplegest is a research accelerator based on public information. It helps an experienced investor review evidence faster, challenge a thesis, and identify the next falsifiable datapoint. The analyst or portfolio manager retains responsibility for source verification, modelling, valuation, and the investment decision.