Executive Summary

Table of Contents

• Product Overview
• Value Proposition
• Core Problem
• Key Differentiators
• Technology Stack
• Current Status
• Quick Links

Product Overview

Praetor (branded as Kairos for hackathon submissions) is a production-grade SaaS platform that enables non-technical users to generate detailed, actionable software specifications through adaptive AI workflows. The system transforms vague project ideas into comprehensive technical specifications through intelligent question-driven discovery, multi-agent AI coordination, and pattern-based knowledge accumulation.

The platform operates on a 6-phase workflow model that progressively refines user intent into production-ready specifications, complete with Definition of Done criteria, test plans, task graphs, and execution runbooks.

Value Proposition

One-Paragraph Summary: Praetor eliminates the gap between business stakeholders and development teams by using AI agents to conduct structured discovery conversations, automatically extracting technical requirements, generating comprehensive specifications, and producing actionable implementation artifacts—all without requiring users to have technical expertise or development experience.

For Business Stakeholders: Transform your project vision into detailed technical specifications without writing a single line of code or understanding technical jargon.

For Development Teams: Receive specifications that are already structured for implementation, with clear acceptance criteria, test plans, and task breakdowns.

For AI-Assisted Development: Export specifications in formats compatible with Claude Code, Cursor, and other AI coding assistants via MCP (Model Context Protocol) integration.

Core Problem

Traditional software specification suffers from several critical failures:

1. Communication Gap: Business stakeholders struggle to articulate technical requirements; developers struggle to understand business context.

2. Incomplete Discovery: Manual discovery processes miss edge cases, fail to explore alternatives, and produce inconsistent documentation.

3. Static Documentation: Traditional specs become outdated immediately; there's no feedback loop between discovery and specification.

4. No Quality Gates: Specifications are "done" when someone decides they're done, not when measurable completeness criteria are met.

5. Lost Knowledge: Each project starts from scratch; patterns and decisions aren't captured for reuse.

Praetor solves these problems through:

• Adaptive AI questioning that ensures comprehensive coverage
• Ontology tracking that monitors what's been discovered vs. what's missing
• Pattern recognition that applies learned knowledge to new projects
• Mechanical completion gates that prevent premature "done" declarations
• Multi-agent validation that catches inconsistencies and gaps

Key Differentiators

1. Conversational Requirements Discovery

Praetor uses conversational AI to extract structured requirements with confidence scoring:

• Requirements Discovery (Migration 094): Natural conversation with real-time extraction to 7 structured JSONB fields
• Confidence Scoring: Every extracted fact has a 0-1 confidence score; low-confidence items become tasks
• Complete Audit Trail: Every extraction tracked with source message, confidence, and extraction type
• Feature Seeds: Automatically generates priority hints (critical/high/medium/low) from discovery for intelligent feature matching
• Graph-Based Discovery (Legacy): Template-driven questions (~15 questions) stored in graph_nodes still supported
• Hybrid Bridge: Seamlessly maps both formats to unified interface for backwards compatibility

2. Intelligent Answer Prefilling

The system doesn't just ask questions—it predicts answers:

• Persona Matching: Identifies user personas (e.g., "technical founder", "enterprise PM") and auto-fills common answers
• Pattern Application: Applies learned patterns from similar projects
• Confidence Scoring: Each prefill includes confidence scores; low-confidence items become tasks for human review

3. Cost-Optimized Model Routing

LLM calls follow a tiered routing strategy:

Cache → Rules Engine → Tier-1 (Fast/Cheap) → Tier-2 (Powerful/Expensive)

This reduces costs by 60-80% while maintaining quality for complex decisions.

4. Mechanical Completion Gates

AI agents cannot falsely claim "done":

• Definition of Done (DoD): Compiled criteria that must be mechanically verified
• Test Plans: Generated test cases with pass/fail gates
• Task Graphs: DAG-based dependency tracking ensuring all prerequisites complete
• Finish Line Contract: Final convergence loop that patches failures automatically

5. Multi-Agent Orchestration

Multiple specialized agents work in coordination:

• Sequential, parallel, hierarchical, and debate execution strategies
• Checkpoint-based recovery for long-running operations
• Budget enforcement (token limits, cost caps)
• Mistake detection and self-healing

6. Open-Core Architecture

Strategic separation of framework (open-source) vs. intelligence (proprietary):

• Open: Event system, schemas, UI components, reference implementations
• Closed: Feature signals, completion thresholds, prefill intelligence, pattern catalogs

7. MCP Execution Concierge

Export specifications directly to AI coding assistants:

• Claude Code integration via MCP server
• Cursor compatibility
• Guided execution with progress tracking

8. Extensive Knowledge Base

Domain expertise through standardized entities, architectural patterns, and AI-generated features:

• Industry Standards: Schema.org, NAICS, FHIR, HR Open, Ed-Fi, SALI, ISO 20022, RESO (thousands of standardized entities)
• Architectural Patterns: Cloud-agnostic and cloud-specific patterns (AWS, GCP, Azure)
• AI-Generated Features: Complete with pattern linkages, question pools, and cloud awareness
• Intelligent Feature Matching: Structured signal extraction (30-40% accuracy improvement over keyword matching)
• Cloud-Aware Recommendations: Suggests cloud-specific patterns (AWS, GCP, Azure) or agnostic alternatives

9. Enterprise-Grade Artifact System

Permissions, workflows, templates, integrations, and schemas as first-class versioned artifacts:

• RBAC + ABAC permissions with policy evaluation
• BPMN workflow parsing and validation
• Document templates with variable extraction and rendering
• Curated + contract-first integrations
• Canonical domain objects with schema registry
• Gap detection with task synthesis
• Export bundling with readiness gating

Technology Stack

Backend

Frontend

AI/LLM

Infrastructure

Current Status

Production-Ready Components

In Progress

Metrics

• Codebase Size: ~70,000+ lines of TypeScript (~10,000 from Engine Framework)
• Database Migrations: 211 migrations (latest: Engine Framework handoffs, extractions, checkpoints)
• Test Coverage: 46 E2E test suites + 460 engine framework tests + 79 requirements discovery tests
• API Endpoints: 97+ routes
• Database Tables: 73+ tables (3 new engine tables: eng_handoffs, eng_extractions_log, eng_state_checkpoints)
• Agents: Multiple specialized agents working in coordination
• Workflows: 10 Mastra workflows
• Knowledge Base:
  • Extensive standardized entity library (multiple industry standards)
  • Architectural patterns (cloud-agnostic and cloud-specific)
  • AI-generated features with pattern-feature mappings verified
• Feature Matching: Structured signal extraction (30-40% accuracy improvement)
• Engine Framework:
  • 2 production engines (Project Discovery, Exploration)
  • 2 stub engines (Business Profile, Ideation)
  • 7 conversation modes (Extract, Explore, Inform, Validate, Refine, Persuade, Deliberate)
  • 16+ conversation techniques (HUMINT, Motivational Interviewing, Cognitive Interview)
  • 7 completeness models (obligation, decision-readiness, maturity, estimability, coverage, profile, saturation)

Quick Links

Last Updated: February 23, 2026