WorkChi

The AI-Native OS for Running Your Business: Build your AI dream team. Virtual agents handle marketing, sales, and customer work for SMEs.

Comment

Inspiration

I was spending $3,000+ per month on 6 different SaaS tools to run my business. But the real problem wasn't the cost - it was that none of them talked to each other.

My CRM couldn't see my support tickets. My calendar didn't know about my deals. My email tool had no idea who my best customers were. Every tool was a silo.

The AI problem: When I tried to add AI features, each tool's AI could only see its own data - maybe 15-20% of the full picture. The best AI in the world is useless if it only sees 20% of your business. You can try Zapier integrations—but businesses don't want to share sensitive data across dozens of third-party suppliers via API. Each integration increases attack surface, compliance scope, and vendor risk.

The security problem: I needed enterprise-grade security with extremely granular permissions—not just "who can see what," but "which LLM providers can access which data types." Full data isolation between workspaces. Complete control over which AI models touch customer PII vs. general business data. The big SaaS players don't give you that control.

I built WorkChi to fix both problems: unified data for smarter AI, with the security controls enterprises demand.

What it does

WorkChi gives startups an AI workforce - not just chatbots, but intelligent agents that see your whole business:

AI Agents for Startups:

  1. Scout - AI lead discovery agent. Automatically finds, scores, and qualifies leads. Generates personalized outreach. (Will be powered by CloudFlare's "pay per crawl" once out of beta)

  2. Virtual SDR - Qualifies inbound leads, sends follow-ups, books meetings. Remembers every interaction via SmartMemory.

  3. Virtual AE - Prepares proposals, handles objections, tracks deal progress. Full context from SDR handoff.

  4. Virtual Marketing Manager - Creates campaigns, writes content, optimizes messaging.

  5. Virtual Customer Success - Monitors customer health, sends check-ins, flags churn risk.

  6. Meeting Agent - Auto-joins meetings when you're late, transcribes in real-time, extracts action items. (integrated with teams, zoom and meet)

  7. Quote Agent - Generates proposals in minutes

  8. Campaign Automation - Multi-channel sequences (Email → SMS → Task) that run autonomously.

  9. NLP Query Agent - Ask anything in plain English: "Show me deals closing this month over $50k"

  10. Meeting Scheduler - Full Calendly replacement, integrated with everything.

The key differentiator: All agents share context via SmartMemory. When Virtual SDR qualifies a lead and hands off to Virtual AE, nothing is lost. Every agent sees the full picture.

How we built it

Built with Claude Code - 87,000+ lines of production TypeScript, developed with AI assistance.

Security-First Architecture:

  1. WorkOS Authentication - Enterprise SSO, RBAC with 34 granular permissions. Not bolted on - built from day one.

  2. Multi-Tenant D1 Isolation - Each workspace gets complete database isolation on Cloudflare D1. Your data never touches another customer's data. Not row-level security - actual database separation.

  3. Model Control - You choose which AI models can access your data. Full audit logging of every AI interaction.

LiquidMetal Raindrop Integration:

  1. SmartMemory - Stores AI context across sessions. Our Virtual Employees remember every customer interaction, pain point, and conversation detail. When you hand off from AI SDR to AI Account Executive, context transfers seamlessly. Namespace isolation: workspace_{id}.

  2. SmartSQL - Natural language database queries. Users can ask "Show me all deals closing this month over $50k" and get instant results. No SQL knowledge required. Works across our 86-table schema.

  3. SmartBuckets - Document storage with RAG. Meeting notes, proposals, and contracts are semantically searchable. AI can find "what did John say about pricing?" across all documents.

Vultr Cloud Inference: All AI reasoning - email drafting, lead scoring, meeting summaries, content generation - powered by Vultr's serverless inference.

We're based in Luxembourg, targeting SMEs in the financial sector—a small market but exceptionally high-value (EU's highest GDP per capita, Europe's largest investment fund hub). No commercial TTS or STT APIs exist for Luxembourgish, so we're building our own on Vultr GPU, leveraging academic foundations like LUX-ASR for speech recognition, YourTTS multilingual models for synthesis, and LuxemBERT for language understanding from the University of Luxembourg. | Enables us to serve an underserved market with no commercial alternatives. Competitive moat through language exclusivity.

Tech Stack:

  • Next.js 15 + React 19 (Server Components)
  • Cloudflare Workers (Edge compute)
  • Cloudflare D1 (Multi-tenant SQLite with full isolation)
  • WorkOS Authentication (Enterprise SSO + RBAC)
  • LiquidMetal Raindrop (SmartMemory + SmartSQL + SmartBuckets)
  • Vultr Cloud Inference (Primary AI) + GPU (Luxembourgish)
  • Drizzle ORM (Type-safe database access)

Challenges we ran into

  1. Multi-tenant isolation - Ensuring workspace data stays isolated in SmartMemory. Solved with namespace-based isolation: workspace_{id}.

  2. AI context window limits - SmartMemory stores full history, but we need to intelligently select relevant context for each query.

  3. SmartSQL schema mapping - Teaching SmartSQL our complex 86-table schema required careful documentation of relationships.

  4. SmartSQL SDK/Documentation Mismatch - The SmartSQL docs show client.executeQuery.execute(), but this API doesn't exist in Node.js SDK v0.6.43. We solved this by implementing a circuit breaker pattern with HTTP API fallback:

    • 8-second timeout per request
    • Circuit opens after 3 failures
    • Auto-fallback to Cloudflare Workers AI

Accomplishments we're proud of

  • 4 Beta Customers - Real users signedup for testing the platform 2nd week of January
  • $150K Potential Enterprise Deal - Demo scheduled January 2026
  • 10+ AI Agents working together with shared memory
  • 87,000+ lines of code - production-ready we use Claude Code, Mini Max M2 and GLM 4.6 spent ~$3000 in tokens
  • Multi-provider AI (Cloudflare + Vultr + Mistral+ Cerebras fallback) we auto choose the best LLM based on the task and data sensitivity

What we learned

An all in one system has dozens of pages, learning curve and seems "busy":

  1. Started with traditional CRM UI (clicks and forms)
  2. Added chat interface (ask questions, get data)
  3. Ended up building "Vibe Biz OS" mode - complete NLP-based business operations that allows running a business all via chat.

Gamification drives employee engagement: We added XP, streaks, goals, and leaderboards for team members. Turns out, making work feel like progress keeps people engaged. Sales reps compete on leaderboards, streaks encourage daily logins, and XP rewards completing tasks. It's not just a CRM - it's a system people actually want to use.

What's next for WorkChi

  1. Vibe Biz OS - Improve NLP-based way to run a business. No screens, just conversation. "Prepare for my meeting tomorrow" and agents do the rest with minimal UI widgets in sidepanels.

  2. Enterprise Enhancement - $150K opportunity suggest this is not just for SMEs, Enterprises want our AI layer on top of Salesforce/SAP. We're building integrations.

  3. AI Agent Studio - Let startups create custom agents with their own personalities, workflows, and integrations.

Built With

  • claude
  • cloudflare
  • cloudflare-d1
  • cloudflare-workers
  • d1
  • liquidmetal-raindrop
  • llama3
  • nextjs
  • react
  • smartbuckets
  • smartmemory
  • smartsql
  • tailwindcss
  • typescript
  • vultr-cloud-inference
  • vultr-gpu
  • workos
Share this project:

Updates