36views

MemoryVerse AI '26

Overview

Live app: upload a PDF → auto-classified (category, skills, summary) via one Gemini call.
Search tab: natural-language queries ("show my AI projects") ranked by semantic similarity, not keywords.
Browse tab: all docs auto-sorted by category. Timeline tab: chronological view from extracted dates.
Relationships: /api/relationships/build — one batched Gemini call finds cross-document connections (not per-pair, avoids O(n²) calls).
Cold start note: backend is on Render free tier — first request after inactivity can take 30–60s. Not a bug.
Design choice: single Gemini call per doc (classify + extract + tag skills together) — keeps latency low, stays in free-tier limits.
Design choice: pure-Python cosine-similarity vector search instead of ChromaDB — its compiled dependency broke on Windows mid-build; swapped deliberately for reliability under deadline. Fine at student-scale (tens of docs), documented as a tradeoff.
Originally Railway-hosted; migrated to Render + Vercel after Railway's trial expired mid-build.
Local setup, full API docs, and architecture diagram are in the repo README.
Limitations stated upfront: no auth (single-user demo scope), relationship engine is one LLM pass not a formal knowledge graph, dates depend on being present in source text.
Solo build, ~48 hours: backend, AI integration, frontend, deployment all done end-to-end.