Dec 2025 to present · React · Supabase · Deno · TypeScript · PostgreSQL
Riven
A student OS, with AI study tools on web and iOS.
- lower cost per request
- 67%
- fewer input tokens
- 70%
- strict citation accuracy
- 61% to 100%
- Deno Edge Functions
- 39
What it is
Riven is a full-stack study platform. It brings AI study tools, collaborative groups, spaced repetition and gamified progress into one place. It runs on React 19 with native iOS support through Capacitor, backed by 39 Deno Edge Functions and a 48-table PostgreSQL schema.
Answers with citations
Students need to trust where an answer came from. I shipped a hybrid RAG pipeline: pgvector and PostgreSQL full-text search, fused with reciprocal rank fusion, using Gemini embeddings. Every answer cites a page or a timestamp.
On a 49-question evaluation set, compared with full-context prompting, the pipeline cut input tokens by 70% and cost per request by 67%. Strict citation accuracy rose from 61% to 100%.
Measuring every call
I instrumented every LLM call with per-request tracing of tokens, latency, time to first token and cost. I also built an LLM-judged evaluation harness for correctness, citations and refusals. It caught a bug in refusal matching.
An AI tutor for any material
The tutor lets students study from notes, PDFs, audio and YouTube transcripts, and generates flashcards and exams automatically. It runs on a multi-model pipeline (Deepgram, GPT, Claude) with SSE streaming, persistent jobs and per-user rate limits.