Trellis

AI visibility and integrity platform, captures every prompt, response, and behavioral signal so organizations can trust how AI-assisted work actually gets done.

Comment

Inspiration

Students are using AI. Employees are using AI. But the people responsible for learning outcomes, instructors, managers, and organizations have zero visibility into how that AI was used. They see the final submission. They don't see the process. We built Trellis to close that gap.

What It Does Trellis is an AI visibility and integrity platform. It captures the entire AI-assisted workflow, every prompt, every response, every behavioral signal, and makes it reviewable.

Behavioral Flag Engine: real-time detection of prompt injection, ghostwriting, answer seeking, and disengagement patterns

Authorship Attribution: tracks whether each word was typed, dictated, pasted, or AI-assisted

Proof-of-Learning Chain: SHA-256 hash chain over every AI interaction, anchored to Solana devnet for tamper-evident audit trails

Multi-Model Support: works across Claude, GPT-4o-mini, and Gemini simultaneously

AI Assignment Generator: upload a PDF or URL and generate a structured assignment with guardrails in under 30 seconds Sensitive Content Detection: warns users before accidentally sending API keys or credentials to external AI services

How We Built It Frontend: Next.js 16, React 19, TypeScript, Tailwind CSS v4 Backend: Supabase (PostgreSQL + Row Level Security), Next.js API routes

AI: Direct HTTP calls to Anthropic, OpenAI, and Google APIs, no SDKs, no vendor lock-in. Graceful fallback to scripted mode when no API keys are configured.

Integrity: Custom SHA-256 proof chain with Solana devnet anchoring

Deployment: Vercel with auto-deploy from GitHub

Challenges Getting multi-provider AI to feel seamless was harder than expected, each provider has a different request/response shape, error behavior, and timeout profile. Building a unified adapter that degrades gracefully (rather than breaking the learner experience) required careful design.

The behavioral flag engine also required balancing sensitivity, too aggressive, and it flags normal learning behavior as too lenient, and it misses what matters.

What We Learned Process visibility is a fundamentally different product from content moderation. We're not trying to catch cheaters, we're trying to give organizations the context they need to trust AI-assisted work. That reframe changed how we designed almost every feature.

What's Next

  • Google Classroom and LTI 1.3 integration for Canvas/Blackboard/Moodle
  • FERPA compliance documentation
  • Organization-level analytics dashboards
  • Real-time instructor alerts

Built With

Share this project:

Updates