Project Helper - Your AI-Powered Business Operating System

Project Helper is an AI-first business management platform that replaces forms and menus with natural conversations. Powered by Google's Gemini AI with real-time function calling, it lets you manage clients, track projects, generate documents, and send emails—all by simply talking. Say "Invoice Acme Corp for $5,000"—the AI finds the client, picks the template, fills the data, and generates the PDF. With voice interaction, semantic search infrastructure, and 30+ intelligent tools, it's not just software—it's a conversational operating system for your business.

Built with Serverpod 3.2.0 (Dart backend), Flutter (cross-platform frontend), Google Gemini AI (with function calling), PostgreSQL + pgvector (semantic search), and Redis (caching).

Inspiration

Modern business software is stuck in the past. Small business owners and freelancers waste hours clicking through menus, filling repetitive forms, and switching between dozens of tools. CRMs feel like complicated databases requiring training manuals. Project management tools are glorified spreadsheets. Document generation means copying and pasting data into templates. Email marketing requires technical skills most people don't have.

We asked ourselves: What if your software just understood what you needed?

What if you could say "I need an invoice for yesterday's client" and the AI would:

Remember which client you met yesterday
Find the right invoice template from your library
Extract amounts and dates from your conversation
Generate a professional PDF instantly
Even send it via email—if you ask

This is software that works the way humans think: through conversation, not clicks.

The Breakthrough Moment: When Google released Gemini with function calling capabilities, we realized AI could finally move beyond chatbots that just generate text. It could actually do things—search databases, create records, generate documents, send emails, update projects. This was the missing piece to build truly conversational software.

Project Helper was born from a simple belief: In 2025 and beyond, the best interface isn't forms and buttons—it's your voice.

What it does

Project Helper is an all-in-one business management platform where you control everything through natural conversation with an AI assistant. Instead of clicking through menus and filling forms, you just talk—and the AI handles the rest.

🤖 AI Assistant with Real Superpowers (30+ Tools)

Unlike chatbots that just generate text, our AI assistant has 30+ intelligent tools that perform real actions in your business:

Partner (CRM) Management:

"Who are my active clients?" → AI searches and lists them
"Add a new contact: John Smith at Acme Corp, john@acme.com" → AI creates the partner record
"When did I last contact Acme?" → AI checks the database and tells you
"Show me all my leads" → AI filters by status

Project Management:

"What projects are due this month?" → AI queries upcoming deadlines
"Create a project called 'Website Redesign' for Acme Corp, starting June 1st" → AI creates it with proper links
"What's the status of the cloud migration?" → AI searches projects and gives you details
"Mark the mobile app project as completed" → AI updates the database

Document Generation (The Magic):

"I need an invoice" → AI shows you visual HTML previews of all invoice templates as cards
You click one → AI asks for details
"It's for $5,000, due June 30th, for Acme Corp" → AI extracts these values from conversation
AI pre-fills the template with smart data → You click "Generate" → PDF downloads

Email Management:

"Email Acme Corp about the project status" → AI finds Acme's email, shows template options
You pick "Project Update" template → AI pre-fills variables from project data
Preview opens with substituted values → You click "Send" → Email sent via SMTP, logged to database

Custom Emails (Ad-Hoc):

"Send email to john@example.com saying 'Thanks for the meeting'" → AI sends immediately without templates
Perfect for quick replies and one-off messages

Smart Context Awareness:

AI knows your data: "You have 15 partners, 8 active projects, 3 deadlines this week"
AI learns your naming: Uses your custom status names
AI remembers conversations: Extracts data from earlier messages

🎙️ 2. Voice-First Experience

Full Voice Interaction:

Speak your requests using speech-to-text (works on all platforms)
AI speaks back using text-to-speech (natural voice responses)
Hands-free operation perfect for mobile, accessibility, multitasking
"Hey AI, what's on my plate today?" → Hear your schedule read aloud

📊 3. Comprehensive CRM (Partner Management)

Rich Contact Management:

Store clients, leads, vendors, contractors
Track status (active, inactive, lead, potential, archived)
Automatic "last contact" tracking (updates when you send emails)
Semantic search ready with pgvector embeddings (1536-dimension vectors stored)
Custom fields via JSON (add any data you need)

Smart Search:

Search by name, email, company
Filter by status
AI-powered natural language queries: "Find my inactive clients from 2024"

📁 4. Project Management with Relationships

Full Project Lifecycle:

Name, description, status tracking
Start dates, end dates, duration calculations
Partner association: Link projects to clients
Sub-projects: Create parent-child hierarchies
Custom statuses: Define your own workflow stages
Default email templates per project

Dashboard Insights:

Projects by status (active, completed, pending, on-hold)
Upcoming deadlines (next 30 days highlighted)
Recent projects with partner info
Filter by partner, status, date ranges

📄 5. Intelligent Document Templates

Professional Document Generation:

Template types: Invoices, contracts, proposals, SOWs, reports, letters
System templates: Pre-built professional templates for everyone
User templates: Create custom templates for your brand
HTML-based: Full styling control with CSS
Variable system: {{clientName}}, {{amount}}, {{date}} placeholders
Client-side PDF generation: Fast, secure, no server upload

AI Template Generation:

"Create a modern invoice template with a blue header" → Gemini generates HTML
"Update this contract to add a termination clause" → AI modifies existing templates
System prompts ensure print-friendly, professional, semantic HTML5

Visual Template Picker:

When AI suggests templates, you see live HTML previews as interactive cards
Click to select (no typing template names)
See before you generate

📧 6. Professional Email System

Rich Email Templates:

Subject templates with variables: Follow-up with {{clientName}}
Plain text and HTML body support
Template types: welcome, followup, notification, reminder, newsletter
Variable preview before sending (see what's missing)

Three Sending Modes:

Template to Partner (Branded Communications):
- Auto-fills partner data (name, email, company)
- Links to projects (include project variables)
- Updates partner's "last contact" timestamp
- Logs to email history
- Supports CC and attachments
Bulk Template Send (Marketing):
- Send same template to multiple partners
- Individual variable substitution per recipient
- Track success/failure per partner
Custom Email (Quick Messages):
- Direct email without template
- Subject + body (text or HTML)
- Instant sending (no preview)
- Perfect for ad-hoc replies

Email History & Logging:

Every email logged to database
Track: recipient, subject, template used, variables, timestamp, success/failure
Filter by partner (all emails to client)
Filter by template (all uses of specific template)
Recent email activity on dashboard
Email activity chart (last 6 months trend)

Production-Ready SMTP:

Configurable SMTP server (SSL/TLS support)
HTML alternatives (plain text fallback)
Attachment support
Test email feature (verify before sending to clients)

📈 7. Real-Time Dashboard & Analytics

Key Metrics at a Glance:

Projects: Total, active, completed, by status, upcoming deadlines
Partners: Total, active, by status
Templates: Document and email template counts
Emails: Sent this month, sent all-time, 6-month activity trend

Recent Activity:

Last 5 projects (with partner info)
Last 5 partners (by last contact date)
Last 5 emails sent (with status badges)

Data Visualization Ready:

Status distribution data (for charts)
Email activity timeline (6-month breakdown)
Deadline tracking (next 30 days)

🔐 8. Enterprise Authentication & User Management

Secure Authentication:

Email/password authentication via Serverpod Auth module
JWT token-based sessions (secure, stateless)
Token refresh endpoint (stay logged in)
Beautiful HTML verification emails
Password reset system
15-minute code expiration

Two-Level User Profiles:

Basic Profile: Username, email, full name, profile image
Additional Info: Custom text field + unlimited JSON data for user settings

Multi-User Support:

Each user has isolated data (row-level security)
All partners, projects, templates scoped to user
Automatic data filtering in queries
Team-ready architecture

How we built it

We built Project Helper using cutting-edge technologies designed for rapid development and scalability:

Backend: We chose Serverpod 3.2.0, a modern Dart-based server framework that eliminates boilerplate through automatic code generation. Define data models in .spy.yaml files, run one command, and get type-safe Dart classes, database migrations, and a complete client SDK automatically generated. This saved us weeks of manual coding.

Database: PostgreSQL with pgvector extension stores all business data plus 1536-dimension vector embeddings for future semantic search capabilities. Redis provides caching for performance.

AI Integration: The heart of the system is Google's Gemini AI with function calling (tool use). We crafted detailed system prompts that train the AI to:

Always fetch real data using tools (never hallucinate)
Chain multiple tool calls intelligently (search client → get project → generate document)
Present results conversationally, not as raw JSON dumps
Understand business context dynamically (knows YOUR partners, YOUR projects, YOUR templates)

The AI receives 30+ function declarations defining available tools (search_partners, create_project, send_email, etc.). When you ask a question, Gemini decides which tools to call, we execute them server-side, and the AI formulates a natural response—all streamed token-by-token for instant feedback.

Frontend: Flutter gives us one codebase that runs on web, iOS, Android, and desktop. We implemented:

Material 3 design with light/dark themes
Speech-to-text and text-to-speech for full voice interaction
Responsive layouts (mobile, tablet, desktop breakpoints)
Real-time streaming chat interface
Visual template pickers with live HTML previews
Client-side PDF generation (no server upload needed)

Architecture: We organized the app into feature modules (ai_chat, partners, projects, documents, emails), making the codebase easy to navigate and extend. Serverpod's built-in authentication handles JWT tokens, email verification, and password resets. Everything is production-ready from day one.

Challenges we ran into

Making AI Actually Intelligent: Gemini's function calling is powerful but tricky. The AI needs to understand when to use tools vs. when to just answer, how to chain multiple tool calls (search client → get project → generate document), and how to present results conversationally instead of dumping raw JSON. We spent countless hours refining system prompts, testing edge cases, and teaching the AI to always fetch real data rather than hallucinate. The breakthrough came when we implemented "forced tool use"—the AI must always check the database, never assume.

Streaming + Tool Calls = Chaos: Gemini streams responses token-by-token for that instant feedback feel. But tool calls interrupt the stream mid-sentence, creating jarring experiences like "I found 3 partn[TOOL CALL]ers...". We solved this by detecting tool call patterns, buffering text, showing "AI is working..." notifications, and resuming smooth streaming after tools complete. The result is seamless even with complex tool chains.

Visual Selection in Text Conversations: Users selecting templates by typing names ("Use invoice_v2") felt clunky. We wanted visual selection with live previews, but chat interfaces aren't designed for this. Our solution: when AI calls show_template_picker, the frontend displays a full-screen modal with gorgeous HTML preview cards. User clicks, selection flows back as a regular message. Conversational flow meets visual interaction.

The Two-Email Problem: Users need both professional templated emails (for clients, with branding and tracking) AND quick ad-hoc emails (replies, casual messages, no overhead). Most systems force one approach. We built two separate pathways that the AI intelligently routes between based on context. "Email Acme Corp" triggers template mode. "Reply saying thanks" sends instantly. Best of both worlds.

Extracting Structure from Conversation: When a user says "Invoice for $5,000 due June 30", how does the AI know these are template variables to extract? We trained the AI to recognize patterns and pass structured JSON ({amount: "5000", due_date: "2026-06-30"}) to the frontend. Forms pre-fill like magic. Users feel like the software reads their mind.

Cross-Platform Voice: Speech-to-text works differently on every platform (web, iOS, Android, desktop). Flutter's plugin has quirks. We wrapped everything with error handling, graceful degradation (text input always available), and clear UI feedback. Tested across browsers and devices until voice worked reliably everywhere.

Performance at Scale: Gemini has 2M token context, but sending full partner/project lists every turn would be slow and expensive. We implemented smart caching (5-minute TTL on system prompts), summary-only data in prompts (counts, not full records), and on-demand fetching (AI tools grab details only when needed). Result: sub-second responses even with hundreds of records.

Preventing Broken Emails: Users would send emails forgetting required variables, resulting in "Hello {{name}}, your invoice is {{amount}}" reaching clients. Embarrassing. We added template parsing to find all {{variable}} placeholders, check which are missing, show warnings in preview, and require confirmation. No broken emails escape.

Accomplishments that we're proud of

We Built AI That Actually Works: Most AI demos are smoke and mirrors. Ours is production-ready. The integration of 30+ Gemini function tools with real database operations, streaming responses, and intelligent context switching is a genuine technical achievement. The AI doesn't just talk—it does things.

Conversational Magic: The document generation workflow genuinely feels like the future. Natural language request → visual template picker with live HTML previews → AI-extracted variables from conversation → one-click PDF. It's delightful. Users are amazed when the AI pre-fills invoice amounts and dates from their casual conversation.

Voice-First Accessibility: Full voice interaction (speech input + voice output) in a business app is rare. We made it work seamlessly across web, mobile, and desktop. The accessibility implications are huge—blind users, motor impairments, people who speak faster than they type. This isn't a gimmick; it's genuinely useful.

The Two-Email System: Elegantly solving the "professional vs. ad-hoc" email problem is a UX triumph. Most business software forces one workflow; we adapted to how users actually work. The AI intelligently routes between template mode (branded, tracked, professional) and custom mode (quick, instant, casual) based on context. It just works.

One Codebase, All Platforms: Flutter gives us web, iOS, Android, and desktop from a single codebase. Same features, same UI, same business logic. The productivity multiplier is enormous. We can ship to all platforms simultaneously.

Enterprise-Ready from Day One: JWT authentication, row-level security (users can't access each other's data), email verification, password reset, production SMTP configuration, proper secret management. This isn't a prototype—it's deployable to real businesses today.

Future-Proof Architecture: We built semantic search infrastructure (pgvector with 1536-dimension embeddings) that isn't even used yet. When we're ready for "Find clients similar to Acme Corp" or "Show projects related to cloud migration," the foundation is already there. Forward-thinking.

Polish Everywhere: Loading states, helpful error messages (not "Error 500"), toast notifications, keyboard shortcuts, responsive breakpoints, light/dark themes, markdown rendering in AI responses. We sweated the details that make software feel professional.

What we learned

AI Function Calling is the Future: Traditional software forces users to click buttons in the right order. AI software lets users speak naturally, and the AI decides what to do. This is a paradigm shift. In 5 years, form-based software will feel as outdated as command-line interfaces. The key insight: AI needs guardrails. We learned to craft system prompts that force tool use, prevent hallucination, and maintain conversational tone. Building AI apps is as much prompt engineering as traditional software engineering.

Modern Frameworks are 10x Multipliers: Coming from Node.js and Python/Django, discovering Serverpod was a revelation. Code generation eliminates boilerplate. Type safety catches bugs at compile time. Automatic migrations replace manual SQL scripts. Built-in auth saves weeks of work. We built in weeks what would take months in traditional frameworks. Choose your tools wisely—they matter.

Voice Should Be Default, Not Optional: We added voice as an experiment. It became a core feature. Why it matters: accessibility (blind users, motor impairments), convenience (hands-free operation), speed (speaking is faster than typing), and naturalness (feels like talking to a person). Voice interaction changes how people think about software. It should be default, not an afterthought.

Users Want Smart Defaults: We initially built tons of customization options. Users ignored them. What worked: AI suggests templates based on context, pre-fills variables from conversation, automatically tracks last contact dates. Make the common case effortless. Advanced users will find settings if needed. Don't make users configure things the software can infer.

Perception Matters More Than Reality: When AI responses loaded all at once, 2 seconds felt slow. When we added token-by-token streaming, 10-second responses felt fast. Streaming creates a sense of progress. Same lesson for visual previews: showing template names in a list is boring; showing live HTML previews is delightful. Use rich media and progressive feedback everywhere.

Testing AI is Like Testing UX: Unit tests for deterministic code are easy. Testing AI responses is hard because they vary. Our approach: test the tools (does search_partners return correct results?), test system prompts (does AI call the right tools?), and manual conversation testing (qualitative evaluation). AI testing is more art than science.

Database Design for AI is Hybrid: Traditional databases are structured (foreign keys, normalized tables). AI-ready databases add vector columns for semantic search, JSON fields for flexible attributes, and full-text indexes. This hybrid approach (structured + flexible + semantic) is the future of data architecture.

Developer Experience Compounds: We obsessed over clear error messages, hot reload, type safety, consistent naming, and good documentation. Result: adding features is a joy, not a chore. When development is smooth, you build faster and make fewer mistakes. DX isn't luxury—it's productivity.

What's next for Project Helper

Semantic Search Activation: We've already built the infrastructure—1536-dimension vector embeddings stored on every partner and project. Next step: populate them with Gemini and unlock "Find clients similar to Acme Corp" and "Show projects related to cloud migration." Meaning-based search, not just keywords.

Mobile Apps: Flutter gives us 90% code reuse for iOS and Android. Deploy to app stores with push notifications (deadline reminders, email alerts), offline mode (work without internet, sync later), and mobile-optimized voice (hands-free operation everywhere). Business management from your pocket.

Team Collaboration: Transform from personal tool to team platform. Multi-user workspaces with shared partners/projects/templates, role-based permissions (admin, editor, viewer), activity feeds, and task assignment. The AI becomes your team's assistant, not just yours.

Advanced Analytics: Interactive charts, trend analysis, revenue forecasting, and AI-powered insights. Ask "Why did sales drop in Q2?" and get data-driven explanations. Ask "Which partners are at risk?" and get predictive analytics. Turn data into decisions.

Ecosystem Integrations: Connect with Google Calendar/Outlook (sync deadlines), Google Drive/Dropbox (file attachments), Slack/Teams (notifications), QuickBooks/Xero (financial sync), and communication tools (SMS, WhatsApp). Project Helper becomes the central hub for all business tools.

Custom AI Training: Fine-tune Gemini on organization-specific data. Learn industry terminology (legal, medical, real estate). Build custom tools for proprietary systems. The AI adapts completely to how YOU work.

Enterprise Scale: SSO/SAML authentication, two-factor auth, comprehensive audit logs, data encryption, SOC 2 compliance, multi-tenant architecture, horizontal scaling. Ready for Fortune 500 companies.

Global Expansion: Multi-language UI and AI conversations, region-specific templates, GDPR compliance, international date/currency formatting. Take this worldwide.

The Vision

Project Helper is a proof of concept for a bigger idea: Software should adapt to humans, not the other way around.

For decades, we've trained employees to use software. We've created user manuals, training sessions, onboarding programs. We've accepted that software is hard to use.

But why? Why should a business owner learn where buttons are? Why memorize required fields? Why click through 5 screens to send an email?

The AI revolution changes everything:

Imagine talking to your software like a colleague: "Hey, I need an invoice for yesterday's meeting."

Your software understands context: "Yesterday you met with John from Acme Corp. Here's an invoice for $5,000 as discussed."

It learns preferences: "I've used your preferred template and included your standard payment terms."

It proactively helps: "Reminder: 3 projects have deadlines this week. Want to review them?"

This is not science fiction. This is Project Helper today.

The next phase is scaling this vision to every business operation:

AI Accountant: "How much did I make last quarter?" → Instant financial analysis
AI HR Manager: "Find candidates for senior developer role" → Resume screening, interview scheduling
AI Sales Assistant: "Draft proposal for Acme's cloud migration" → Generates from past projects
AI Customer Support: "Handle customer complaints" → Reads emails, drafts responses, escalates when needed

The goal: A business operating system where AI handles routine tasks, and humans focus on creativity, strategy, and relationships.

Project Helper is the first step. The future is conversational. The future is intelligent. The future is now.

Built with passion by developers who believe software should serve humans, not the other way around.

"The best way to predict the future is to invent it." — Alan Kay