Built to be deployed, not just demonstrated
A single ASP.NET Core MVC application over one SQL Server database. Deliberately boring infrastructure, so the interesting part can be the intelligence layer.
Platform
- .NET 8 (LTS)
- ASP.NET Core MVC
- Razor views
- Dependency injection throughout
Data
- SQL Server
- Entity Framework Core 8
- Code-first migrations
- Indexed, normalised schema
Security
- ASP.NET Core Identity
- Role-based authorisation policies
- Anti-forgery on every unsafe verb
- Audit logging
Front end
- Bootstrap 5
- Custom design system
- Chart.js for analytics
- Font Awesome
Real time
- SignalR notification hub
- Per-role delivery groups
- Persisted before delivery
- Graceful degradation without JS
Processing
- Hosted background worker
- Durable job queue
- No long work on request threads
- Structured logging via Serilog
Architecture decisions worth stating
One deployable application, not twelve projects
Separation of concerns is achieved through folders, interfaces and services inside a single project. A small team has to understand, debug and deploy this during a hackathon; a twelve-project solution would buy architectural purity at the cost of every practical quality that matters here.
The AI layer sits behind an interface
IAiService and ISkillExtractionService define the contract. The shipped provider is a gazetteer-based extractor that runs fully offline with no API key, so the platform is never dependent on an external service being reachable. Swapping in a hosted model means registering a different implementation and changing nothing else.
No hardcoded intelligence
Every weight, threshold and normalisation constant lives in strongly typed configuration. Demand scoring, drift classification, gap severity and curriculum alignment can all be retuned per deployment without a code change.
Charts are never hardcoded in JavaScript
Every chart reads a JSON payload the server produced from live queries. The client-side code knows how to render a chart; it never knows what the values are.
Derived data is recomputed, never hand-edited
Demand scores, drift results, district gaps and curriculum alignment are all outputs. They are replaced wholesale on each pipeline run, which means they cannot silently diverge from the evidence.
Project structure
SkillSync/ ├── Controllers/ thin, no business logic ├── Models/ domain entities by area ├── ViewModels/ presentation shapes ├── Data/ DbContext, migrations, seed ├── Services/ application services ├── Interfaces/ service contracts ├── AI/ extraction, normalisation │ └── Providers/ pluggable AI implementations ├── Analytics/ the intelligence engines │ ├── SkillDemandService │ ├── SkillDriftService │ ├── SkillGapService │ ├── CurriculumAlignmentService │ ├── DistrictIntelligenceService │ └── RecommendationService ├── Infrastructure/ config, cross-cutting ├── Middleware/ global exception handling ├── Hubs/ SignalR ├── Views/ Razor └── wwwroot/ css, js, assets SkillSync.Tests/ unit tests
Request flow
Controller → Service → EF Core → SQL Server
Controller → AI Service → IAiService → Provider
Controller → Queue → Background Worker → Engines
Controllers stay thin. Long-running intelligence work is queued rather than executed on the request thread, so no HTTP request ever waits on a full pipeline run.