Browse⌘K
tsfeatured.ts
tsxtrellis.tsx
tsxtrellis.tsx

Trellis

2025live
$2,500 · Vibe Coding Challenge · Springfield, MO

AI workflow visibility platform — captures every prompt, response, and behavioral signal so evaluators can review the process behind AI-assisted work, not just the final output.

Trellis

Story

Students use ChatGPT; instructors have no idea if learning happened. Trellis flips the model: instead of trying to block AI, it captures the full process — prompts, responses, behavioral signals, paste detection, engagement scores — so instructors can review the learning, not just the final submission.

Direct HTTP calls to three model providers (Claude, GPT-4o-mini, Gemini), no SDK lock-in, minimal bundle. Learners pick their model mid-session; every interaction gets proof-chain hashed. A scripted fallback mode runs full demos with zero API keys for offline pitching.

What I built

  • 1Built the multi-provider AI adapter with proof-chain hashing across Claude, OpenAI, and Gemini
  • 2Designed the behavioral flag engine for real-time prompt-injection / disengagement / help-seeking detection
  • 3Implemented the AI assignment generator: PDF or URL → structured assignment with guardrails in seconds

Tech

Next.js 16React 19TypeScriptTailwind v4SupabasePostgreSQLClaudeGPT-4o-miniGeminiSolanaZodVercel

Links

Live siteDevpost

Up Next

● open to SWE roles·trellis.tsx·UTF-8·github: …
·resume·contact··