CrossCheck

AI-powered knowledge audit tool — conversational, adaptive, brutally honest about your gaps.

CrossCheck lets you upload study notes and runs you through a live Q&A session that escalates in pressure as you progress, then produces a structured breakdown of exactly what you know and what you don't. Built as a full-stack solo project in React 19 + TypeScript, powered by Gemini 2.5 Flash, NVIDIA NIM vision models, and Supabase.

---

Live demo

https://youtu.be/QDo0nn68ewU

crosscheck-j4vf.onrender.com

Upload notes	Live audit	Diagram questions	Knowledge report
PDF, image, or plain text	Conversational examiner, adaptive pressure	Visual challenges from your diagrams	Strong / Weak / Needs Revisit per topic

---

How it works

Most study tools quiz you passively. CrossCheck interrogates you the way an examiner would — conversationally, with follow-ups, escalating difficulty, and no hand-holding.

1. Upload notes — PDF, plain text, or images. Client-side extraction via pdfjs-dist + Gemini Vision.
2. Topic extraction — Gemini maps key concepts from the notes automatically.
3. Adaptive audit session — The AI examiner shifts mode as the session progresses:
   • Friend (0–25%) → warm, relaxed warm-up
   • Tutor (25–50%) → guided understanding
   • Instructor (50–75%) → precise, no soft nudges
   • Examiner (75–100%) → rigorous, minimal reactions
4. Diagram questions — Select pages with diagrams in your PDF. NVIDIA NIM (llama-3.2-11b-vision-instruct) generates 3 visual challenge questions per page — function, relationship, and process angles. Text labels are redacted from the image using pdfjs exact text coordinates before being shown to the student. Questions are injected every 4th AI turn with a blur-to-reveal image card and fullscreen lightbox.
5. "I don't know" — One tap delivers the correct answer, marks the topic weak, and continues with a simpler follow-up.
6. Overtime mode — If weak or untested topics remain when time expires, the session continues in Examiner mode focused only on those gaps.
7. Knowledge report — Every topic scored Strong / Weak / Needs Revisit with supporting evidence. Weak topics expand to show specific concepts to revisit.
8. Session persistence — All sessions are saved to Supabase. Resume any incomplete session from the home screen exactly where you left off.
9. Personality mode — Upload exported chat logs and Gemini extracts speech patterns, phrases, and humor style to build a custom AI persona that quizzes you in someone's voice. PIN + email gated.

---

Engineering highlights

Streaming-first AI pipeline

All Gemini calls use generateContentStream rather than generateContent. The SDK throws on structured JSON output that gets truncated before the closing brace — streaming and consuming chunks manually sidesteps the validation entirely.

Resilient JSON recovery

Gemini's structured output can return malformed JSON under token pressure. A 6-stage safeParseJSON pipeline handles it without crashing:

1. Direct parse
2. Clean (trailing commas, unquoted keys)
3. Slice to last valid closing brace
4. Repair truncated arrays/objects
5. Truncation fallback — extract whatever completed before the cutoff
6. Regex field extraction as last resort

Diagram question pipeline

PDF pages are rendered to <canvas> via pdfjs-dist. Text positions are extracted from the PDF's text layer (exact pixel coordinates, no OCR guessing) and painted over on a second canvas to produce a label-free copy for the student. The original labeled image is sent to NVIDIA NIM via a Supabase Edge Function proxy (to avoid CORS) which generates targeted visual questions. Three questions per page are generated with rotating angles (function → relationship → process) so no two questions test the same dimension of understanding.

Session persistence and resume

Every session is upserted to Supabase after each turn. The resume flow restores full state — topics, messages, elapsed time, note content, and any in-memory diagram questions. Base64 image data is stripped before DB writes so storage stays lean.

Personality extraction

Conversation logs (WhatsApp, iMessage, Discord) are analyzed by Gemini to extract a PersonalityStyle schema: speech patterns, phrase vocabulary, emoji frequency, humor type, reaction style. The result is injected as a system instruction, transforming the examiner's tone in the Friend phase only.

Client-side document processing

All document processing runs in the browser — no backend file upload, no server storage. PDFs are processed page-by-page, images are passed as inline base64, DOCX files via mammoth. The NVIDIA API call is the only one that routes through a backend (Supabase Edge Function) to avoid CORS.

---

Stack

Layer	Technology
Frontend	React 19, TypeScript, Vite
AI — session	Google Gemini 2.5 Flash (@google/genai)
AI — diagrams	NVIDIA NIM llama-3.2-11b-vision-instruct
Auth + DB	Supabase (email auth, PostgreSQL, Edge Functions)
PDF parsing	pdfjs-dist (client-side render + text extraction)
Styling	Inline styles + CSS variables (5 themes: light, dark, pink, ocean, emerald)

---

Local setup

npm install
npm run dev

Create .env.local:

VITE_GEMINI_API_KEY_1=your_gemini_key
VITE_SUPABASE_URL=your_supabase_url
VITE_SUPABASE_ANON_KEY=your_supabase_anon_key
VITE_NVIDIA_API_KEY=your_nvidia_nim_key   # from build.nvidia.com — required for diagram questions

Supabase falls back to a no-op mock if the URL is not set. The NVIDIA key is optional — diagram question generation fails gracefully (pages are skipped with a logged error) if absent.

Supabase Edge Function

The NVIDIA vision calls are proxied through a Supabase Edge Function (supabase/functions/nvidia-vision/) to avoid CORS. Deploy it once:

supabase functions deploy nvidia-vision

---

Deploy

A render.yaml is included for one-click Render static site deployment with correct MIME types and SPA routing. For other hosts:

Build command: npm run build
Publish directory: dist

Add the .env.local variables as environment variables in your hosting dashboard.

---

Built by push — also check out oracleai.live.