TradeGuard AI

A multi-agent intelligence platform for retail traders.

"Every retail trader is alone in the room. TradeGuard AI puts an AI council next to them."

What It Does

Retail traders lose money not because markets are hard — but because they act on emotion, incomplete information, and hype while institutions operate with teams of analysts, quant models, and real-time data feeds.

TradeGuard AI closes that gap with a swarm of specialized AI agents, each owning one intelligence domain, running in parallel, and converging into a single coordinated verdict.

The Agent Swarm

User Action / Ticker Query
         │
    ┌────▼─────────┐
    │  AI Router   │   lib/ai/router.ts + Redis cache
    └────┬─────────┘
         │
   ┌─────┴──────────────────────────────────┐
   │                                        │
   ▼                                        ▼
Azure OpenAI GPT-4o                  Claude 3.5 Sonnet
(deep_research → synthesis)          (news_analysis · journal_reflection
   │                                  morning_brief)
   │  fallback ────────────────────────▶   │
   │                                        │
   ▼                                        ▼
Grok Beta                            Perplexity Sonar-large
(hype_detection)                     (news_research · citations)
   │                                        │
   └──────────────────┬─────────────────────┘
                      │
              ┌───────▼────────┐
              │  Redis Cache   │   shared memory across agents
              └───────┬────────┘
                      │
              ┌───────▼──────────────┐
              │  Synthesis Agent     │   Azure GPT-4o aggregates
              │  (Azure GPT-4o)      │   all agent outputs
              └───────┬──────────────┘
                      │
              ┌───────▼──────────────┐
              │ Azure Content Safety │   journal guardian (non-blocking)
              └───────┬──────────────┘
                      │
                 Final Output

Agent responsibilities

Agent	Model	Task type	Role
Research Agent	Azure OpenAI GPT-4o	`deep_research`	Final synthesis — coordinates all agent outputs
News Agent	Claude 3.5 Sonnet	`news_analysis`	Impact analysis, retail trap detection
Journal Agent	Claude 3.5 Sonnet	`journal_reflection`	Behavioral coaching, emotion tagging
Brief Agent	Claude 3.5 Sonnet	`morning_brief`	Daily pre-market intelligence
Social Agent	Grok Beta	`hype_detection`	X/social sentiment, FOMO detection
News Research Agent	Perplexity Sonar	`news_research`	Verified facts with source citations
Safety Guardian	Azure Content Safety	—	Screens all journal entries before DB write

Key swarm properties:

Agents run via Promise.all — parallel, not sequential
Redis cache layer allows agents to reuse each other's results
Azure GPT-4o → Claude fallback keeps the swarm alive if Azure is unavailable
Azure GPT-4o is the synthesis coordinator — waits for all agents, produces the final verdict

Azure Integration

Two Azure services are load-bearing in the swarm.

Azure OpenAI GPT-4o — Synthesis Agent

lib/ai/router.ts routes deep_research tasks to Azure GPT-4o first, with automatic Claude fallback:

case 'deep_research':
  try {
    responseText = await callAzureOpenAI(prompt, options)
  } catch (azureErr) {
    console.warn('Azure OpenAI unavailable, falling back to Claude:', azureErr)
    responseText = await callClaude(prompt, options)
  }
  break

lib/services/research.service.ts routes the final synthesis step through deep_research, meaning every Research Terminal query closes with Azure GPT-4o synthesizing all three parallel agent outputs into one verdict.

Azure Content Safety — Journal Guardian

Every journal entry is screened before it touches the database (app/api/journal/route.ts):

const safety = await checkContentSafety(rawText)   // non-blocking

await db.journalEntry.create({
  data: { ...entry, riskFlag: safety.flagged }
})

Categories checked: Hate · SelfHarm · Sexual · Violence
Severity scale: 0 (safe) → 6 (high)
riskFlag: true written to DB → warning badge shown in UI
Journal always saves regardless — Content Safety never blocks the write

Modules

Intelligence Feed

Real-time market event analysis, delivered before you can react.

Polygon API ingests live news on a Vercel cron schedule
Claude analyzes each item: what happened, what it means for retail traders, retailTrap flag
Output: sentiment score, impact level (LOW / MEDIUM / HIGH / CRITICAL)
Live-pushed to all connected clients via Pusher WebSocket

Research Terminal

Institution-grade parallel research on any ticker in under 10 seconds.

T=0s  Yahoo Finance   → live price, volume, EMA, RSI
T=0s  Perplexity      → verified news (last 24h) with source URLs
T=1s  Claude ×3       → technical read · news impact · retail trap (parallel)
T=3s  Azure GPT-4o    → final synthesis of all three agent outputs
T=4s  Result rendered

Supports: US equities, NSE/BSE (India), crypto, forex, ETFs.

Trading Journal

The only trading journal with a behavioral AI coach built in.

Free-text entry → Azure Content Safety screens it first (non-blocking)
Claude identifies emotion tag: FOMO · REVENGE · DISCIPLINED · FEARFUL · PANIC · GREED
Structured AI response: pattern name, likely mistake, one action for tomorrow
PsychProfile builds over time: dominant biases, streak tracking, weekly insight
riskFlag stored in DB — visible warning badge if Content Safety flags the entry

Tech Stack

Layer	Technology
Frontend	Next.js 14, React 18, TypeScript, Tailwind CSS, Framer Motion
UI Components	shadcn/ui, Lucide Icons, Lightweight Charts
AI — Synthesis	Azure OpenAI GPT-4o
AI — Safety	Azure Content Safety
AI — Analysis	Anthropic Claude 3.5 Sonnet
AI — Social	Grok Beta (xAI)
AI — News	Perplexity Sonar-large
Database	PostgreSQL via Supabase, Prisma ORM
Cache	Upstash Redis
Real-time	Pusher WebSocket
Auth	Clerk
Data Providers	Polygon, Yahoo Finance, Alpaca, FMP, StockTwits
Deployment	Vercel (crons + edge)

Project Structure

├── app/
│   ├── (dashboard)/
│   │   ├── feed/           # Intelligence Feed
│   │   ├── research/       # Research Terminal
│   │   ├── journal/        # Trading Journal
│   │   ├── watchlist/
│   │   └── settings/
│   └── api/
│       ├── feed/
│       ├── research/[symbol]/
│       ├── journal/
│       └── cron/           # Vercel background jobs
├── lib/
│   ├── ai/
│   │   ├── router.ts               # Agent dispatcher
│   │   ├── azure-openai.ts         # Azure GPT-4o client
│   │   ├── azure-content-safety.ts # Content Safety client
│   │   ├── claude.ts
│   │   ├── grok.ts
│   │   ├── perplexity.ts
│   │   └── prompts.ts
│   ├── services/
│   │   ├── research.service.ts     # Parallel ticker analysis
│   │   ├── feed.service.ts
│   │   ├── journal.service.ts
│   │   └── brief.service.ts
│   └── data/               # Market data providers
└── prisma/
    └── schema.prisma

Setup

1. Install dependencies

pnpm install

2. Configure environment variables

cp .env.local.example .env.local

Fill in all values. The example file has inline comments for every variable explaining where to find it. Required to get started: DATABASE_URL, DIRECT_URL, NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY, CLERK_SECRET_KEY, ANTHROPIC_API_KEY.

Azure variables (AZURE_OPENAI_*, AZURE_CONTENT_SAFETY_*) are optional locally — both clients degrade gracefully when not configured (Content Safety returns flagged: false, Azure OpenAI falls back to Claude).

3. Generate Prisma client and run migrations

pnpm prisma generate
pnpm prisma migrate dev

4. Run the development server

pnpm dev

Background Jobs (Vercel Crons)

Route	Schedule	What it does
`/api/cron/morning-brief`	Daily 9:00 AM ET (Mon–Fri)	AI-generated pre-market brief
`/api/cron/news-ingestion`	Every 15 min, 9 AM–5 PM ET	Ingests and analyzes latest market news
`/api/cron/trending-scan`	Every 30 min	Scans StockTwits for unusual ticker activity
`/api/cron/cleanup`	Daily 1:00 AM	Expires feed events older than 7 days
`/api/cron/regime-refresh`	Daily 2:30 AM (Mon–Fri)	Refreshes ^NSEI and ^GSPC regime snapshots
`/api/cron/trader-models`	Daily 1:00 AM	Recomputes TraderModel for active users
`/api/cron/mind-directives`	Daily 3:00 AM (Mon–Fri)	Generates Mind Directives for active users

All cron routes require a Bearer token matching CRON_SECRET. Generate one with openssl rand -hex 32 and set it in both .env.local and Vercel dashboard.

ULTRA — V7 Mind Engine (current)

Phase 2 adds a proprietary behavioral intelligence layer that fuses market regime with each trader's own verified statistics to produce a personalized daily directive.

Architecture

                   ┌─────────────────────────────────────────┐
                   │           Mind Engine (V7)              │
                   │                                         │
 Market Data ──►  │  Gaussian HMM (regime-service/Python)   │
                   │  ↓ BULL_TREND | BEAR_TREND | CHOP | CRISIS │
                   │                                         │
 Trade History ►  │  TraderModel (pure TS, zero AI)         │
                   │  ↓ winRate, expectancyR, statsBySetup   │
                   │  ↓ 4 behavioral flags (n≥5 threshold)  │
                   │                                         │
 Journals ──────► │  Mind Directive (AI synthesis)          │
                   │  ↓ headline + reads + EV + oneRule     │
                   └─────────────────────────────────────────┘
                            │
                   ┌────────▼────────┐
                   │  Live Copilot   │  ← fused with TraderModel flags
                   │  + Directive EV │    and statsAfterLoss
                   └─────────────────┘

New in V7

Regime Classifier — Gaussian HMM on 5y daily returns, 4-state labels, 6h Redis cache, heuristic fallback. Python microservice in regime-service/.
TraderModel — pure TypeScript stat engine. Computes winRate, expectancyR, profitFactor, statsBySetup/Regime/Hour, statsAfterLoss/Win, convictionCalib, and 4 behavioral flags. Calibrates at ≥20 trades AND ≥10 journals.
Mind Directive — daily AI synthesis of regime + TraderModel → JSON directive with headline, market read, personal read, EV assessment, greenlight/avoid setups, size guidance, and ONE imperative rule. PRO-gated; FREE users see blurred preview + upgrade CTA.
Copilot × Mind Fusion — behavioral agent receives top 2 flags + statsAfterLoss + today's EV for hyper-personalized real-time warnings.
Trade Log (/trades) — R-multiple tracking, setup tags, conviction stars, close modal with live P&L preview.
Mind Page (/mind) — three-column layout: directive + history, full TraderModelCard, regime panels.

Behavioral Flags

Flag	Trigger
`OVERSIZE_AFTER_WINS`	Avg position size after 2+ consecutive wins > 1.3× baseline (n≥5)
`REVENGE_FAST_REENTRY`	≥3 same-setup re-entries within 15 min of a loss
`LATE_DAY_LEAK`	Last-hour expectancyR < overall expectancyR − 0.3R (n≥8)
`CONVICTION_INVERTED`	Win rate at 5★ < win rate at ≤2★ (both n≥5)

Honesty Guarantees (V7)

Every AI output labeled with actual model used
Every statistic shows sample size: 58% (n=31)
TraderModel shows CALIBRATING when <20 trades or <10 journals
Approximations labeled (approx) — regime from heuristic fallback
App never places trades; disclaimer on every analysis surface

Roadmap

V1 — Current

Intelligence Feed with retail trap detection
Research Terminal — 4-agent parallel analysis
Trading Journal with behavioral AI coach
Azure OpenAI GPT-4o — synthesis agent
Azure Content Safety — journal guardian with riskFlag
Real-time Pusher feed
PsychProfile — emotion tracking over time

V7 — Mind Engine (shipped)

Gaussian HMM regime classifier (Python microservice)
TraderModel — 4 behavioral flags, zero AI
Mind Directive — daily AI directive, PRO-gated
Copilot × Mind Fusion — flags injected into behavioral agent
Trade Log with R-multiple tracking
/mind page — three-column layout
Extended demo seed — 30 closed trades, all flags firing

V8 — Execution Intelligence (next)

Broker webhook integration (Zerodha Kite, IBKR)
Pre-trade checklist gate — blocks entry until all rules verified
Emergency intervention — full-screen modal when psychState = TILT
Position sizing calculator respecting TraderModel size guidance
Streak-based regime alerts (3+ consecutive regime changes)

License

MIT / Proprietary

Name		Name	Last commit message	Last commit date
Latest commit History 43 Commits
app		app
components		components
lib		lib
prisma		prisma
public		public
regime-service		regime-service
scripts		scripts
types		types
.env.local.example		.env.local.example
.env.production.example		.env.production.example
.eslintrc.json		.eslintrc.json
.gitignore		.gitignore
LAUNCH.md		LAUNCH.md
README.md		README.md
SUBMISSION.md		SUBMISSION.md
components.json		components.json
middleware-disabled.ts		middleware-disabled.ts
middleware.ts		middleware.ts
next.config.js		next.config.js
next.config.mjs		next.config.mjs
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
postcss.config.mjs		postcss.config.mjs
tailwind.config.ts		tailwind.config.ts
tsconfig.json		tsconfig.json
vercel.json		vercel.json

Folders and files

Latest commit

History

Repository files navigation

TradeGuard AI

What It Does

The Agent Swarm

Agent responsibilities

Azure Integration

Azure OpenAI GPT-4o — Synthesis Agent

Azure Content Safety — Journal Guardian

Modules

Intelligence Feed

Research Terminal

Trading Journal

Tech Stack

Project Structure

Setup

1. Install dependencies

2. Configure environment variables

3. Generate Prisma client and run migrations

4. Run the development server

Background Jobs (Vercel Crons)

ULTRA — V7 Mind Engine (current)

Architecture

New in V7

Behavioral Flags

Honesty Guarantees (V7)

Roadmap

V1 — Current

V7 — Mind Engine (shipped)

V8 — Execution Intelligence (next)

License

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages