Retail investors are forced to bounce between 10+ tools — Yahoo Finance, TradingView, SEC EDGAR, DART, FRED, FMP, broker apps, YouTube earnings calls, news scrapers — to do what one Bloomberg seat does in one window. The information asymmetry costs them real money.
**ATLAS Terminal closes that gap as a single desktop-first interface** that fuses market data, fundamental research, valuation, technicals, macro, filings, news, video transcript analysis, and a printable institutional report — without farming financial computation out to an LLM.
---
## Core Principle
> **LLMs handle text. Python handles numbers.**
Every valuation number, every ratio, every Monte Carlo path is computed deterministically in Python with `pandas`, `scipy`, and `numpy`. Gemini is reserved strictly for qualitative work — MD&A summarisation, 10-K risk extraction, transcript summarisation, news translation, copilot dialogue. This separation keeps the math auditable and the token bill in check.
---
## What's New — Video Transcript Workbench
**Just shipped.** A complete pipeline that turns any video into structured research:
1.**Submit** a YouTube URL, direct media URL, or local upload
2.**Extract** — prefer existing subtitles via `yt-dlp`, fall back to local `faster-whisper` STT
3.**Analyse** — Gemini distils summary · keywords · topics · sentiment · intent in one JSON pass
4.**Persist** — SQLite FTS5 (or PostgreSQL `tsvector`) makes every transcript searchable
5.**Translate** — optional Korean translation on demand
Built so an earnings call, a CEO interview, or a sell-side YouTube deep-dive can become a structured note inside the terminal in a single round trip — never leaving the research workflow.
> Drop in a screenshot of the running `/transcripts` page at `docs/media/atlas-transcripts.png` to display it here.
- **Per-feature fallback chains** — every data domain has an explicit primary→secondary source (`yfinance → yahooquery` for prices, `FMP → yahooquery → yfinance` for historical ratios, `yahooquery → yfinance` for DCF inputs). Failures degrade gracefully, never surface raw exceptions to the UI.
- **One file, one responsibility** — services are capped at ~300 lines, routers stay thin, business logic stays out of UI components.
- **Hybrid LLM separation** — all numeric work in Python (`dcf_engine.py`, `monte_carlo.py`, `financial_metrics.py`, `risk_metrics.py`), all text work routed through `gemini_service.py` with 60-second 429 back-off and `smart_chunk()` token compression
- **Per-feature fallback chains** documented in `claude.md` §2.3 — each endpoint has an explicit primary→secondary source order
- **Async background jobs without a broker** — transcript ingestion uses `asyncio.create_task` + status polling, no Celery/Redis dependency for single-instance deployment
- **Multi-jurisdiction filings** — ticker-suffix routing (`/filings`) auto-picks SEC, DART (`.KS` / `.KQ`), or EDINET (`.T`) so research flow doesn't break across markets
- **DB-agnostic** — `unified_repo.py` routes every call to `aiosqlite` or `asyncpg` based on `DATABASE_URL`; FTS5 ↔ `tsvector` swap is transparent
I built ATLAS Terminal because the asymmetry between what a Bloomberg seat shows a fund analyst and what a retail investor sees on Yahoo Finance is enormous — and it's a tooling problem, not a data problem. The raw data is public. The synthesis is what's missing.
- design a coherent product spanning **market data, valuation modelling, technical analysis, macro, multi-jurisdiction filings, and video research** under one shell
- enforce a **hard architectural rule** (LLMs for text, Python for numbers) and defend it across 37 services and 22 routers
- ship the full stack — frontend, backend, database layer, CI, encrypted credential vault, async background jobs — solo
- keep the work auditable: deterministic financial math, multi-source fallbacks, no LLM-priced calculations, no leaked exceptions to the UI
