GoBot — AI Prospect Research Agent

GOBOT — AI PROSPECT RESEARCH & OUTREACH AGENT Built by A Development | Internal Tool → Client-Ready Product OVERVIEW GoBot is an AI-powered Telegram bot that researches businesses, audits their web security, generates case study packages, and drafts outreach emails — all from a single command. What used to take 2-3 hours of manual research per prospect now takes 45 seconds. GoBot runs 24/7 on a local machine, uses 9 LLM providers with automatic failover, and has generated over 50 prospect packages since launch. THE PROBLEM Prospecting for web development and AI services is research-intensive. Each qualified lead requires: — Website audit (speed, SEO, security, mobile, content) — Competitive landscape review — Contact discovery — Personalized outreach email — Internal pitch strategy notes Doing this manually for 10 prospects per day = 20-30 hours per week. Unsustainable. WHAT WE BUILT Architecture Telegram → relay.ts → Skill Router → LLM Provider Chain ↓ ↓ MCP Tools Memory (Supabase) (Playwright) ↓ ↓ Conversation History Templates → HTML Files → Disk + Telegram + Dashboard The 9-Provider LLM Failover Chain GoBot never goes offline due to a single provider outage. It cycles through: 1. Ollama (local, qwen3:8b) — zero cost, instant 2. Mistral (devstral-latest) 3. Qwen Coder (OpenRouter, free tier) 4. Groq (llama-3.3-70b-versatile) 5. Gemini (gemini-2.0-flash) 6. OpenRouter (Meta Llama 3.3 70B, free) 7. NVIDIA Inference (Llama 3.3 70B) 8. Cerebras (llama3.1-8b, ultra-fast) 9. SambaNova (Meta-Llama-3.1-405B — largest available) Combined cost: ~$0/month for typical usage (free tiers + local Ollama). The 6-File Prospect Package Each /prospect command generates a complete outreach package in ~45 seconds: 01-email-to-[contact]-[business].html Personalized cold email referencing specific findings from the audit. Uses the business's actual pain points, not generic templates. 02-[name]-case-study-[date].html Client-facing case study showing what A Development would build for them. Structured like a proposal with problem → solution → expected outcomes. 03-[name]-security-audit-[date].html Full security audit: SSL, headers, known CVEs, exposed endpoints, CMS version, third-party script risks. Graded A-F. contact-info.txt Discovered contact name, email, phone, LinkedIn, and social handles. Internal/solutions.html (not shared with prospect) Technical recommendations and implementation notes for internal use. Internal/pitch-answers.html (not shared with prospect) Anticipated objections and prepared responses for the sales call. Skills (10 Commands) /prospect <url> Full 6-file package /prospect audit <url> Quick security audit only /prospect search <industry> <location> Find prospects matching criteria /prospect email "<business>" Draft outreach from existing package /prospect pipeline View all prospects sorted by pipeline stage /prospect band <url> Band/venue booking outreach variant /prospect pitch <industry> Industry-specific proposal template /remember <fact> Store a fact in persistent memory /goals Show active goals and progress /help List all commands + MCP status Memory System (Supabase) GoBot remembers everything across sessions: — Facts stored via /remember (e.g., "Preferred contact time: mornings") — All conversations logged to messages table — Goals tracked with completion status — Semantic search via embeddings for context retrieval Web Dashboard (localhost:3737) Local web UI for reviewing prospect packages, downloading files, checking LLM provider health, and managing memory. Accessible from any device on the local network. RESULTS Efficiency: — Manual prospect research: 2-3 hours per prospect — GoBot: 45 seconds per prospect — Speed improvement: 160-240x — Cost per prospect: ~$0 (free tier LLMs + local Ollama) Output Quality: — Security grades match manual Qualys/SSL Labs results within 1 grade — Cold email open rates higher than template-based outreach — Contact discovery accuracy: ~80% (vs ~60% manual) Reliability: — 9-provider failover chain = 99.9%+ availability — Runs on local hardware, no cloud dependency — Automatic retry logic with provider health monitoring TECH STACK Runtime: Node.js, TypeScript AI: Claude API, Ollama (local), Groq, Gemini, Mistral, OpenRouter, NVIDIA, Cerebras, SambaNova Browser Automation: Playwright (via MCP server) Database: Supabase (PostgreSQL) — memory + conversation history Frontend: HTML5, CSS3 (dark theme design system) Transport: Telegram Bot API Infrastructure: Local machine + Cloudflare Tunnel (gobot.adevelopment.dev) DESIGN SYSTEM GoBot output uses A Development's dark design system: — Background: #0a0a0a / Surface: #141414 / Borders: #2a2a2a — Font: Inter (Google Fonts CDN) — 20-color accent pool (rotating per prospect for visual distinction) — Grade colors: A=green, B=blue, C=orange, D=red, F=dark red — Animated gradients, card layouts, responsive WHAT'S NEXT GoBot is being packaged as a white-label product for marketing agencies and business development teams. The core research pipeline (MCP + Playwright + LLM chain) is reusable across any vertical. Current integrations in development: — /evlv skill: EV charging infrastructure research for LAW Ventures — Auto-push to ADEV dashboard via webhook — Fine-tuned outreach model trained on high-performing emails

GOBOT TECHNICAL ARCHITECTURE CORE COMPONENTS 1. relay.ts (Telegram → LLM bridge) Receives Telegram messages, routes to skill handlers, manages LLM provider selection and failover. Stateless message processing with persistent Supabase memory for cross-session context. 2. Skill Router Pattern-matched command dispatcher. Each /skill is a self-contained module with its own prompt template, MCP tool set, and output formatter. Skills can chain sub-commands (e.g., /prospect audit runs inside /prospect). 3. LLM Provider Chain Ordered list of providers with health monitoring. On failure, automatically advances to next provider. Provider selection can be overridden per-skill (e.g., code generation always tries devstral first). 4. MCP Server (Playwright) Model Context Protocol server exposing browser automation tools to the LLM. The agent can navigate websites, extract structured data, take screenshots, and fill forms — all through natural language tool calls. Key MCP tools: — browser_navigate(url) — browser_snapshot() → accessibility tree — browser_click(element) — browser_type(element, text) — browser_evaluate(script) → DOM queries 5. Template Engine HTML templates per output type (case study, security audit, email, pitch). Variables injected at render time from LLM-extracted data. Each template uses the ADEV dark design system for visual consistency. 6. Memory Layer (Supabase) Two tables: — memory: key-value facts with embeddings for semantic search — messages: full conversation history with role/content/timestamp Embeddings generated via OpenAI ada-002 (or local alternative). Semantic search surfaces relevant past context for each new message. SECURITY MODEL — Telegram webhook validated via secret token — No external ports exposed (Cloudflare Tunnel only) — LLM API keys stored in .env, never logged — Prospect data stored locally, never uploaded to cloud — Dashboard access restricted to local network DEPLOYMENT — Node.js process managed by PM2 (auto-restart on crash) — Cloudflare Tunnel: gobot.adevelopment.dev → localhost:3737 — Git-based updates (git pull + pm2 restart) — No Docker required (bare metal for low latency Ollama)