VibeBullish is an agentic quant trading platform

Generative AI works the data floor as labeler, narrator and risk flagger. A LightGBM ensemble ranks the universe, and a deterministic verdict mapper turns each prediction into typed decision rows in action_decisions — the one currency every surface consumes. Paper books trade those rows under pre-registered rules. Nothing trades live.

Model cohorts

tradeable_v2 + tradeable_mid_large

Universe sweeps

4×/day

plus a 30-min stale refresh in market hours

Paper books

5 active

4 equity + BTC Trend · public standings route, 2026-09-29

Live books

paper only

Five repositories

One shared PostgreSQL database, five separately deployed repositories.

Repository	Stack	Role	Deploy
vibebullish-backend	Go / Gin	Action Generation Engine, Auto-Pilot V2 paper books, catalyst scanner, server-driven UI feeds, WebSocket hub	Railway, auto on push to `main`
vibebullish-data-pipeline	Python	LightGBM training, on-demand `predict_one` and gated batch scoring, news and SEC forward loops, the `dq-daily` data-quality gate	Railway, auto on push to `main`
vibebullish-web	Next.js 16	Stock feed, ticker details, search, Auto-Pilot, Scoreboard, Lab, SDUI renderer	Vercel, manual CLI deploy
vibebullish-ios	SwiftUI / XcodeGen	Native client that renders the same server-driven feeds	App Store
vibebullish-data-pipeline-dashboard	Static HTML	Ops console: LLM usage and system health	Vercel, auto on merge to `main`

This docs site also ships by manual Vercel CLI deploy. Details: Deployment.

How services connect

1 · Market data

Polygon (Massive)

Real-time stocks WebSocket (minute aggregates + trades) and REST

Finnhub

Analyst data and company news

SEC filings

8-K and Form 4 text via the Massive filings API, read by the pipeline

Nasdaq Trader halts RSS

Trading-halt events

2 · Backend scanner + feature substrate

Scanner

Runs continuously with market-aware scheduling and queues tickers for the engine

ticker_features_daily

Per-(ticker, date) ML feature cache the pipeline reads for training and inference

3 · Pipeline inference

predict_one

Per ticker, on demand. Every call passes one concurrency limit and a circuit breaker.

score_batch

Requested by the backend at 16:45 ET, only after the dq-daily gate is ready

4 · Action Generation Engine

Predict → score → map

Ensemble prediction, then MapVerdict. Display-only risk and quality scores are stamped afterwards. No LLM in this path.

Five trigger modes

scheduled_batch, scanner_signal, on_demand, catalyst_event, price_gap_refresh

action_decisions

The engine's only output table

5 · Consumers

Feeds

Server-driven UI on web and iOS, with feed_update WebSocket events

Paper books

Auto-Pilot V2 enters and exits paper positions from decision rows

Tripwires + push

Price-target tripwires and notifications

Ops dashboard

LLM usage and system health

Step by step: How It Works. Sources in depth: Data Sources.

Key tables

Table	Writer	Purpose
`action_decisions`	backend	Typed decision rows per ticker and horizon
`ticker_features_daily`	both	Per-(ticker, date) ML feature cache
`daily_bars`	both	Daily OHLCV bars
`ticker_fundamentals`	both	Latest fundamentals per ticker
`books` / `book_positions` / `book_daily_equity`	backend	Paper book ledger, positions and daily equity snapshots
`llm_catalyst_emissions`	backend	LLM-labelled catalysts; input to the risk and quality scores and the catalysts section
`news_articles`	pipeline	Polygon and Finnhub articles, labelled in the pipeline
`lgbm_test_predictions`	pipeline	Out-of-sample test predictions from training runs
`action_resolutions`	backend	Graded outcomes of past decisions

Server-driven UI and the feed plugin system

Server-driven UI

SDUI

The backend returns a ServerDrivenFeedResponse: a list of typed ServerDrivenComponent specs.
Web and iOS map each component type to a native view.
The web renderer’s COMPONENT_MAP holds 42 component types; unknown types show a diagnostic placeholder in development and render nothing in production.

Feed plugins

Go backend

Each feed implements FeedTypePlugin: GetSupportedTypes, GenerateFeed, GetPriority.
The aggregator routes a request by feed type; a new feed is one plugin plus RegisterPlugin.
Public feed plugins: action, autopilot, crypto, search.
Auth is Firebase middleware.

State as of 2026-09-29

What runs today

dated observation

Verdicts are pure quant. The per-decision LLM reviewer was removed on 2026-06-14. No production path calls Opus since 2026-09-14.
LLM roles: labeler, narrator, risk flagger. Model names are vendor aliases.
The V1 signal loop was removed on 2026-09-15.
Two production cohorts: tradeable_v2 and tradeable_mid_large.
Crypto is parked: a data-only market refresh, no verdicts, no LLM calls.
No live orders. A live book needs §19 graduation, a second non-overlapping window of ≥90 trading days, execution-readiness and shadow-live qualification, and signed owner approval.
Engineering shadow books run beside the measured books and are excluded from §19.

Public paper roster

GET /api/autopilot/standings

Book	Label	Status
57	Pure Quant	active
58	Insider	active
59	Analyst Consensus	active
60	Pure Quant · Trailing	active
42	BTC Trend	active
53–56	Previous epoch of the same four strategies	retired 2026-09-16

Next: How It Works · Action Engine · Auto-Pilot (books, §19 criteria) · Data Sources · Deployment

Verified against vibebullish-docs@0a4edb1, vibebullish-backend@07c641de, vibebullish-data-pipeline@ee230ea, vibebullish-web@ce8ec92 on 2026-09-29. Runtime observations are dated inline; a source sha does not vouch for deployed configuration or database contents.