Inspiration
Software repositories contain an enormous amount of engineering intelligence, but most existing code-analysis tools reduce that intelligence to static dashboards, isolated security findings, or a single numerical score.
We wanted to build something different.
YOWON AI was created around a simple question:
Can an AI system actually understand a software project deeply enough to help engineering teams make deployment and architecture decisions?
Instead of treating a repository as a collection of files, YOWON AI treats it as an engineering system. It reconstructs the repository's architecture, dependencies, technology stack, security posture, engineering health, and project DNA before asking AI agents to reason over the evidence.
What YOWON AI Does
YOWON AI is an AI Engineering Intelligence Platform that transforms a Git repository into an evidence-grounded engineering intelligence report.
The pipeline begins with the repository itself:
Repository → Repository Intelligence → Project DNA → Evidence → AI Council → Deterministic Scoring → LLM Reasoning → Decision Intelligence
The platform analyzes the actual codebase before generating its conclusions.
It can identify:
- Architecture and application layers
- Programming languages and frameworks
- Dependencies and technology stack
- Security findings
- Testing and documentation maturity
- Maintainability indicators
- Repository health
- Technical risks
- Engineering recommendations
- Deployment readiness
Multi-Agent Engineering Council
One of the core ideas behind YOWON AI is that software quality should not be judged from a single perspective.
The platform uses a multi-agent engineering council where specialized agents examine different dimensions of the repository.
Forge
Acts as the architecture specialist and analyzes software structure, modularity, dependencies, and engineering design.
Sentinel
Focuses on security posture, vulnerabilities, authentication paths, and security findings.
Guardian
Evaluates operational and engineering risks and identifies factors that could affect reliability.
Visionary
Examines innovation, technology choices, AI capabilities, and technical differentiation.
Showcase
Reviews presentation and developer-facing quality.
Prime
Acts as the chief synthesis agent, bringing the evidence and specialist evaluations together into an executive engineering decision.
The goal is not to make the agents produce independent opinions without evidence. Their outputs are grounded in repository intelligence and deterministic evaluation data.
Repository Intelligence
Before the AI agents reason about a project, YOWON AI builds a structured understanding of the repository.
The system extracts:
- Repository structure
- Source files
- AST information
- Dependencies
- Frameworks
- Languages
- Databases
- AI frameworks
- Vector stores
- CI/CD technologies
- Architecture relationships
- Engineering metrics
- Security findings
- Evidence references
This creates a Repository DNA representation of the project.
The resulting intelligence becomes the evidence layer used throughout the evaluation pipeline.
Evidence-Grounded AI
A major design principle of YOWON AI is:
AI should explain engineering decisions from evidence rather than inventing conclusions.
The LLM receives a structured evidence graph containing repository metadata, architecture information, security findings, specialist evaluations, recommendations, metrics, timeline information, and historical evaluation data.
The qualitative report is then generated dynamically by the LLM.
Different report sections use specialized reasoning personas such as:
- CEO / CTO
- Staff Architect
- Principal Security Engineer
- VP Engineering
- AI Research Lead
- Risk Officer
- Staff Engineer
The system also applies hallucination guards and contradiction checks before the final report is persisted.
Dynamic AI Reports
YOWON AI does not rely on a collection of hardcoded report paragraphs.
The Insight reasoning layer generates structured sections including:
- Executive reasoning
- Architecture analysis
- Security reasoning
- Business implications
- Innovation analysis
- Risk analysis
- Recommendations
The generated narrative is specific to the repository being evaluated.
If the available evidence is insufficient, the system is instructed to acknowledge that limitation rather than fabricate an answer.
Explainability and Citations
Every important AI conclusion should be traceable back to engineering evidence.
YOWON AI supports evidence references containing information such as:
- Finding IDs
- Source agents
- File paths
- Line ranges
- Evidence descriptions
- Confidence information
This allows an engineering reviewer to move from an AI conclusion back toward the underlying repository evidence.
Historical Intelligence
Repositories are not static.
A project can improve after developers fix vulnerabilities, add tests, restructure modules, or improve documentation.
YOWON AI therefore compares completed evaluations across time.
Historical intelligence can identify:
- Score improvements
- Score regressions
- Resolved blockers
- Persistent risks
- Changes in engineering health
- Evolution of deployment readiness
This turns the platform from a one-time scanner into an engineering intelligence system that can understand project progression.
Decision Simulation
YOWON AI also includes a decision simulation layer.
Instead of immediately running another complete evaluation, engineers can explore hypothetical improvements such as:
- Fixing authentication
- Increasing test coverage
- Improving architecture
- Adding CI/CD
The simulator recalculates the expected decision metrics on the frontend so engineers can explore:
"What could happen if we fix this?"
This provides a lightweight way to prioritize engineering work.
Enterprise Intelligence Workspace
The final result is presented as an engineering intelligence workspace rather than a conventional static report.
The workspace includes areas for:
- Executive decision intelligence
- Repository intelligence
- AI Council reasoning
- Security and risk analysis
- Architecture analysis
- Business intelligence
- Innovation analysis
- Recommendations
- Historical intelligence
- Evidence and citations
- Report export
The visual system is designed around a dark, high-density engineering environment with unified workspace cards, subtle particle backgrounds, accent borders, and evidence-focused information hierarchy.
Enterprise Architecture
YOWON AI is designed as an extensible enterprise platform rather than a single analysis script.
The platform includes concepts such as:
- Organizations
- Teams
- Members
- Roles and RBAC
- Repository management
- Connectors
- Secrets management
- Webhooks
- Workflow automation
- Audit logs
- Policy controls
- System health
This allows repository intelligence to become part of a broader engineering operating environment.
Authentication and Security
The platform also includes enterprise authentication capabilities including secure password-reset workflows.
The authentication layer uses mechanisms such as:
- OTP verification
- HMAC-based OTP protection
- JWT reset tokens
- Password history
- Replay protection
- Rate limiting
- Session invalidation
- Audit logging
Security is therefore considered both in the repositories being analyzed and within the platform itself.
What We Learned
Building YOWON AI taught us that simply connecting an LLM to a repository does not create meaningful engineering intelligence.
The difficult part is building the evidence layer before the reasoning layer.
We learned to separate:
Deterministic analysis → Evidence → Agent reasoning → Scoring → LLM synthesis → Explainability
This separation makes the system easier to validate and significantly reduces the risk of an AI model generating unsupported conclusions.
We also learned that multi-agent systems become much more useful when every agent has a clearly defined responsibility and receives the same structured evidence context.
Challenges
Some of the biggest challenges were:
- Understanding large repositories without overwhelming the reasoning model.
- Keeping deterministic engineering metrics separate from subjective LLM reasoning.
- Preventing AI-generated reports from contradicting numerical evaluation results.
- Grounding qualitative conclusions in actual repository evidence.
- Preserving historical evaluation data while evolving the report system.
- Designing a UI capable of presenting large amounts of engineering intelligence without turning into a conventional dashboard.
- Maintaining backward compatibility while continuously evolving the evaluation pipeline.
These challenges shaped the architecture of YOWON AI.
Why YOWON AI Is Different
Traditional code-analysis systems generally answer:
"What problems exist in this repository?"
YOWON AI aims to answer a larger engineering question:
"What does the evidence tell us about this software system, why does it matter, what should we do next, and how confident should we be in that decision?"
That shift—from scanning to engineering decision intelligence—is the core idea behind YOWON AI.
Future Direction
The next stage of YOWON AI focuses on making the engineering intelligence system even more autonomous.
Planned capabilities include:
- Multi-agent debate
- Confidence calibration
- Evidence graph visualization
- Decision-tree reasoning
- Automated report quality validation
- Architecture visualization
- Engineering maturity modeling
- Executive AI Q&A
- Cross-repository intelligence
- Portfolio-level risk intelligence
- Continuous repository monitoring
The long-term goal is to build an AI Operating System for Software Engineering where repositories are continuously understood, engineering risks are surfaced proactively, and teams can make decisions using evidence rather than fragmented tooling.
Built With
- ast
- crewai
- fastapi
- framer
- github
- python
- react
- sqlite
- tailwind
- typescript
Log in or sign up for Devpost to join the conversation.