ConduitAI

ConduitAI is an ASP.NET Core MVC application for real-estate lead intelligence and follow-up. It is an internal CRM covering lead records, timeline history, dashboard metrics, stored AI analysis, and meeting-note processing. AI runs locally through Ollama and is invoked only on explicit user actions.

Scope, architecture, and contribution rules are documented in CLAUDE.md and AGENTS.md.

Features

• Lead management — create, edit, delete, list, and detail views with server-side validation and safe delete confirmation.
• Search and filtering — filter leads by status, source, location, free-text search, and minimum AI lead score.
• Lead timeline — record interactions (calls, emails, meetings, property tours, follow-ups, notes); adding one bumps the lead's last-updated time.
• Stored AI lead analysis — on demand, generate a summary, 0–100 lead score, urgency, buying intent, and recommended next action. Results are persisted, never regenerated on page load.
• Meeting notes assistant — turn raw notes into a structured summary, key facts, risks, and a next action; optionally attach to a lead, which also records a timeline interaction.
• Dashboard — total leads, new leads, high-priority leads, upcoming follow-ups, a recently-updated list, and a follow-up queue.

Engineering Highlights

• Local-first AI integration. Lead analysis and meeting-note extraction run against a local Ollama model over a typed HttpClient, using a structured JSON response contract, tolerant parsing, a single parse-and-repair retry, and graceful degradation when the model is unavailable — partial output is never persisted.
• Defense-in-depth by default. Anti-forgery on every POST, ViewModel binding to prevent over-posting, parameterized EF Core LINQ, Razor output encoding for untrusted user and model text, and loopback-only AI requests so lead data never leaves the machine.
• Clean MVC architecture. Thin controllers over a focused service layer (lead, timeline, dashboard, AI analysis, meeting notes), with EF Core/SQLite persistence and migrations.
• Progressive enhancement with jQuery. Pages render server-side and stay fully usable without heavy client logic; jQuery layers on the explicit, user-triggered AI analysis AJAX call (ai.js) and safe delete/confirmation flows (site.js), plus jQuery Validation Unobtrusive for client-side form checks — and AI output is inserted via .text()/DOM APIs, never as raw HTML.
• Tested behavior. xUnit coverage across the service layer, lead filtering, AI JSON parsing, and AI/meeting-note failure paths.

Technology Stack

• ASP.NET Core MVC (.NET 8) with Razor views
• C#
• Entity Framework Core 8 with SQLite
• HTML / CSS / JavaScript / jQuery (progressive enhancement)
• Ollama running locally at http://localhost:11434

Prerequisites

• .NET SDK 8.0
• Ollama for AI features (the rest of the app works without it)

Getting Started

dotnet restore
dotnet run --project ConduitAI

The app applies EF Core migrations and seeds fictional demo data on first run, so a separate dotnet ef database update is not required. The SQLite database is created at ConduitAI/App_Data/conduitai.db (git-ignored).

Then open the URL printed by dotnet run (by default http://localhost:5229).

Enabling AI features

AI actions (lead analysis, meeting notes) require Ollama running locally with the configured model:

ollama serve

In another terminal:

ollama pull qwen2.5-coder:7b
ollama list

If Ollama is unavailable, configured to a non-local URL, or returns unparseable output, AI actions fail gracefully with a user-facing message and nothing partial is stored. All other CRM features remain fully usable.

Configuration

Ollama settings live in appsettings.json under the Ollama section:

"Ollama": {
  "BaseUrl": "http://localhost:11434",
  "Model": "qwen2.5-coder:7b",
  "TimeoutSeconds": 120
}

appsettings.json holds public-safe defaults. To use a different local model without changing committed files, override with environment variables:

Ollama__Model=qwen2.5-coder:14b dotnet run --project ConduitAI

AI requests are only sent to loopback Ollama URLs such as http://localhost:11434 or http://127.0.0.1:11434. Non-loopback URLs disable AI features instead of sending private lead data off-machine.

No secrets are required to run the app.

Project Structure

ConduitAI/
  Controllers/    Thin controllers (Home, Leads, Interactions, Ai, MeetingNotes)
  Services/       Business logic + interfaces and AI result types
  Models/         EF Core entities and enums
  Data/           AppDbContext, design-time factory, seed initializer
  Migrations/     EF Core migrations
  ViewModels/     Form and display models (prevent over-posting)
  Views/          Razor views and partials
  Helpers/        Display formatting helpers
  wwwroot/        site.css, site.js, ai.js
ConduitAI.Tests/  xUnit tests (EF Core InMemory)

Service layer

• LeadService — lead CRUD and filtering
• TimelineService — interaction history
• DashboardService — dashboard metrics
• AiAnalysisService — generates and stores lead analysis
• MeetingNotesService — processes and stores meeting notes
• OllamaClient — typed HttpClient wrapper for the Ollama API
• AiPromptBuilder / AiResponseParser — structured prompt construction and tolerant JSON parsing

Data model

SQLite tables: Leads, LeadInteractions, LeadAnalyses, MeetingNotes. AI results are stored for auditability rather than overwritten or regenerated automatically. Meeting notes attached to a lead are deleted with that lead; unattached notes are only created through the standalone meeting-notes workflow.

Testing

dotnet test

Tests cover the service layer, lead filtering, AI response parsing, and AI/meeting-note failure paths using the EF Core InMemory provider and hand-rolled fakes.

Security & Privacy

This project is built for a public repository, so security is part of the design:

• No secrets, tokens, passwords, real customer data, or private notes are committed; local databases, logs, and secrets are git-ignored.
• Input is validated server-side with DataAnnotations and service-level checks.
• POST forms use anti-forgery tokens; forms bind to ViewModels, not database entities.
• Persistence uses EF Core LINQ and parameterized queries.
• User input and AI output are treated as untrusted and rendered with Razor's default HTML encoding (never as raw HTML).
• Prompts instruct the model not to invent facts and to avoid legal, financial, or protected-class recommendations.
• Raw notes, emails, phone numbers, and full AI responses are kept out of logs.

Constraints

By design, ConduitAI is free to run and develop and adds no authentication, roles, Docker, microservices, Redis, queues, cloud infrastructure, CI/CD, paid APIs, or hosted LLM services. Seed data is fictional.

License

ConduitAI is licensed under the MIT License. Third-party client libraries in wwwroot/lib include their own license files.