Skip to content

mary4data/marktron

BranchesTags

Repository files navigation

marktron_

An autonomous AI marketing agent that monitors brand visibility in AI search results, detects content gaps, and generates drafts to close them — while you sleep.

What is marktron?

When someone asks ChatGPT, Gemini, or Perplexity "what's the best CRM for revenue operations?", your brand may or may not appear. GEO (Generative Engine Optimization) is the emerging discipline of making sure it does.

marktron is a closed-loop GEO agent built for a single brand (Attio by default). It:

  1. Monitors AI visibility scores daily via the Peec AI API
  2. Detects gaps where competitors outrank you in AI-generated responses
  3. Researches what competitor content is winning those gaps
  4. Generates content drafts to reclaim that ground using Gemini AI
  5. Tracks whether published content actually moves the needle over time
  6. Forecasts projected score improvement from approved actions

Architecture

graph TB
    subgraph Browser["Browser (Next.js App Router)"]
        UI["Dashboard UI\n/dashboard/*"]
        PROBE["GEO Probe\ngpt-4o-mini + Gemini"]
    end

    subgraph Server["Server (Next.js API Routes)"]
        AGENT["/api/agent\nGemini Agentic Loop"]
        CYCLE["/api/run-cycle\nDeterministic Pipeline"]
        REPORT["/api/report\nPDF Generation"]
        DELTA["/api/attio/delta\nForecast Engine"]
        OV["/api/attio/overview\nBrand Landscape"]
    end

    subgraph AI["AI Layer"]
        VERTEX["Vertex AI\nGemini 2.5 Flash\n(content generation)"]
        VERTEX2["Vertex AI\nGemini 2.5 Flash\n(agent function calling)"]
        OPENAI["OpenAI\ngpt-4o-mini\n(GEO Probe)"]
    end

    subgraph External["External APIs"]
        PEEC["Peec AI\nVisibility scores\nPrompts · Gaps"]
        TAVILY["Tavily\nCompetitor source search"]
    end

    subgraph Storage["Google Cloud Firestore"]
        SNAP["visibility_snapshots"]
        DRAFTS["content_drafts"]
        BRANDS["brands"]
        SOURCES["competitor_sources"]
    end

    subgraph Output["Output"]
        PDF["PDF Report\n@react-pdf/renderer"]
        EMAIL["Email Delivery\nResend"]
    end

    UI --> AGENT
    UI --> CYCLE
    UI --> REPORT
    UI --> DELTA
    UI --> OV
    PROBE --> OPENAI
    PROBE --> VERTEX

    AGENT --> PEEC
    AGENT --> TAVILY
    AGENT --> VERTEX2
    AGENT --> Storage

    CYCLE --> PEEC
    CYCLE --> TAVILY
    CYCLE --> VERTEX
    CYCLE --> Storage

    DELTA --> SNAP
    OV --> SNAP
    OV --> DRAFTS

    REPORT --> PDF
    REPORT --> EMAIL
    REPORT --> Storage
Loading

Agent Cycle (How It Works)

The agent runs in two modes:

Autonomous Agent (/dashboard/agent)

Uses Gemini with function calling in an agentic loop. Streams every tool call live to the UI.

Phase 1 — parallel, no LLM (instant):
  ├── get_visibility   → Peec AI: overall score + competitor scores
  └── get_gaps         → Peec AI: topics where competitors win

Phase 2 — no LLM (instant):
  └── save_snapshot    → Firestore: historical record for delta chart

Phase 3 — Gemini function calling, tools run in parallel per turn:
  ├── search_sources × 3   → Tavily: competitor URLs + content snippets
  ├── generate_draft × 3   → Gemini: brand-voice content draft per gap
  └── save_draft × 3       → Firestore: pending drafts for human review

Deterministic Pipeline (/api/run-cycle)

Same steps without the LLM planning layer — faster, no streaming, used for scheduled runs.

Features

Feature Description
GEO Probe Simulate how ChatGPT and Gemini rank your brand vs competitors for any prompt
Overview Brand landscape with live visibility scores, competitor gaps, activity feed
Visibility Gaps Full prompt table from Peec AI with topic, volume, country, priority
Content Queue Review, approve, or reject AI-generated content drafts
Peec Delta Time-series chart: historical scores + 2-day forecast based on approved drafts
Agent Feed Live streaming view of every tool call the agent makes, with collapsible JSON
PDF Report One-page visibility report: scores, gaps, actions, forecast — downloadable or emailed

Tech Stack

Frontend

Layer Technology
Framework Next.js 16 (App Router, TypeScript)
Styling Tailwind CSS v4 with dark: variant, CSS custom properties
Dark mode next-themes (attribute="class", default dark)
Charts Recharts (LineChart, ReferenceArea, ReferenceLine)
Icons Lucide React
Toasts Sonner

Backend

Layer Technology
Runtime Next.js API Routes (Node.js, server components)
Database Google Cloud Firestore (Native mode, custom DB marktron-db)
Auth Application Default Credentials (ADC) — no service account key files
Streaming Server-Sent Events (SSE) via ReadableStream

AI & APIs

Service Purpose
Peec AI Visibility scores, prompt library, gap analysis
Vertex AI / Gemini 2.5 Flash Content draft generation
Vertex AI / Gemini 2.5 Flash Autonomous agent function calling
OpenAI GPT-4o mini GEO Probe — ChatGPT ranking simulation
Tavily Competitor source search (web crawl + ranking)
@react-pdf/renderer Server-side PDF generation

Developer Tooling

Tool Purpose
Entire CLI Session recording — links AI coding sessions to git commits
Firebase Admin SDK Firestore access from server routes

Data Model (Firestore)

brands/{brandSlug}
  ├── name: string
  ├── peec_project_id: string
  ├── is_own: boolean
  └── color: string

visibility_snapshots/{id}
  ├── brand_id: string
  ├── snapshot_date: string          # YYYY-MM-DD, Europe/Berlin timezone
  ├── overall_score: number          # 0–100
  ├── competitor_scores: object      # { "HubSpot": 80, "Pipedrive": 85, ... }
  ├── gap_topics: array              # [{ topic, competitor, gap, your_score, ... }]
  └── created_at: timestamp

content_drafts/{id}
  ├── brand_id: string
  ├── title: string
  ├── body: string
  ├── content_type: string           # blog_post, linkedin, reddit, ...
  ├── target_url: string
  ├── status: "pending" | "approved" | "rejected"
  ├── approved_at: timestamp | null
  └── created_at: timestamp

competitor_sources/{id}
  ├── brand_id, snapshot_id, competitor_name
  ├── source_url, source_type, topic
  ├── tavily_score: number
  └── content_snippet: string

Forecast Model

The delta chart projects 2 days forward using:

projected_score = current_score + (slope × days) + content_boost

where:
  slope         = (last_score - first_score) / (n_days - 1)
  content_boost = approved_count × avg_gap × 0.08
  avg_gap       = mean of gap_topics[].gap from latest snapshot

Forecast lines are dashed in the chart. A vertical ReferenceLine separates historical from projected data.

Environment Variables

# Google Cloud (Firestore + Vertex AI via ADC)
FIREBASE_PROJECT_ID=your-gcp-project-id
FIRESTORE_DATABASE_ID=marktron-db            # optional, defaults to (default)

# Peec AI
PEEC_API_KEY=your-peec-api-key
PEEC_PROJECT_ID=your-peec-project-id

# Tavily (competitor source search)
TAVILY_API_KEY=your-tavily-api-key

# OpenAI (GEO Probe — ChatGPT simulation)
OPENAI_API_KEY=your-openai-api-key

# Resend (PDF email delivery)
RESEND_API_KEY=re_...
REPORT_FROM_EMAIL=reports@yourdomain.com

# App URL (used for absolute links)
NEXT_PUBLIC_APP_URL=https://your-domain.com

Local Development

# 1. Authenticate with Google Cloud (Vertex AI + Firestore via ADC)
gcloud auth application-default login

# 2. Install dependencies
npm install

# 3. Copy and fill environment variables
cp .env.example .env.local

# 4. Start the dev server
npm run dev

# 5. (Optional) Enable Entire session recording
entire enable

Project Structure

src/
├── app/
│   ├── page.tsx                    # Marketing landing page
│   ├── dashboard/
│   │   ├── agent/page.tsx          # Autonomous agent live feed
│   │   ├── probe/page.tsx          # GEO Probe
│   │   ├── overview/page.tsx       # Brand landscape
│   │   ├── gaps/page.tsx           # Visibility gaps table
│   │   ├── queue/page.tsx          # Content queue
│   │   ├── delta/page.tsx          # Peec Delta chart + forecast
│   │   └── settings/page.tsx       # Settings
│   └── api/
│       ├── agent/route.ts          # SSE agentic loop (Gemini function calling)
│       ├── run-cycle/route.ts      # Deterministic pipeline
│       ├── report/route.ts         # PDF generation + email
│       ├── probe/route.ts          # GEO Probe (ChatGPT + Gemini)
│       └── attio/
│           ├── overview/route.ts
│           └── delta/route.ts
├── components/dash/
│   ├── AppLayout.tsx               # Sidebar, header, nav, theme toggle
│   ├── ProbePanel.tsx              # GEO Probe result card
│   ├── RunProbeBar.tsx             # Probe input + engine selector
│   └── EngineBreakdown.tsx         # Per-engine metrics
└── lib/
    ├── firebase.ts                 # Firestore client (ADC)
    ├── peec.ts                     # Peec AI visibility + gap queries
    ├── peec-client.ts              # Peec brand/prompt data
    ├── tavily.ts                   # Competitor source search
    ├── generator.ts                # Gemini content generation (Vertex AI)
    └── report-pdf.tsx              # PDF document component

Live Demo

marktron.lovable.app — frontend prototype built with Lovable

Powered by

Peec AI · Vertex AI · Tavily · Entire · Lovable

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages