MemoryVerse AI '26

MemoryVerse AI — an AI-powered Digital Identity System that turns a student's scattered documents (certificates, projects, resumes, internship letters, GitHub) into one structured, connected, searchable journey — while preserving every original file. Live demo: https://memoryverse-ai-five.vercel.app GitHub: https://github.com/919vaibhavyadav-oss/memoryverse-ai NO LOGIN OR CREDENTIALS NEEDED. The demo is pre-loaded with a sample student journey, so every feature is populated out of the box. (The Groq API key lives server-side — reviewers don't need one.) NOTE: the backend runs on a free tier and sleeps when idle, so the FIRST page load may take ~40–50s to wake up, then it's instant. WHAT TO TRY (about 60 seconds): 1) Dashboard → type "show my AI projects" in the Ask bar → get an instant natural-language answer PLUS the source documents (openable in their original format). This is the core "I never have to search through folders again" moment. 2) Knowledge Graph → hover a skill node (e.g. Python) → its connections across certifications, projects and internships light up. Below it: AI-generated insights + suggested career paths. 3) Add Documents → paste a GitHub profile URL (e.g. github.com/919vaibhavyadav-oss) → it imports every repo as a categorized Project via the GitHub API. 4) Journey → the 2023 → 2026 growth timeline. AI/ML TECHNIQUES (and where in the code): - NLP + LLM reasoning via Groq (llama-3.3-70b): categorization, metadata extraction, relationship insights, and RAG answers (backend/app/services/categorizer.py, retrieval.py, relationships.py) - Embeddings + semantic search: TF-IDF vector space + cosine similarity (backend/app/services/vectorstore.py) - RAG: retrieve top-k chunks by similarity, then Groq synthesizes a grounded answer over only the retrieved context (no hallucinated facts) - Knowledge mapping: an entity graph (documents ↔ skills ↔ orgs) rendered as an interactive force-directed graph DESIGN DECISIONS WORTH NOTING: - Groq handles reasoning; a self-contained pure-NumPy TF-IDF vector store handles semantic search (Groq has no embeddings endpoint). This keeps the project installable anywhere with zero build tools or model downloads, works fully offline, and is swappable for a neural embedder behind the same interface. - Graceful degradation: with no API key the app still runs end-to-end in a rule-based "Heuristic mode" — nothing hard-fails during evaluation. - Original uploaded files are never modified and are always downloadable.