Vidyut
Sriram

Quorum joins a live meeting as a participant: it listens in real time, contributes ideas, speaks back, takes actions like filing a GitHub issue on request, and can drive a shared screen. Built over a weekend on Recall.ai, Deepgram, Claude and ElevenLabs, it attacks the same problem I keep working on in PrismAI.

Python · FastAPI · Claude · Recall.ai · Deepgram · ElevenLabs

Once a day it checks EDGAR for new 8-Ks, marks each one notable or routine with fixed rules on item codes (every verdict names the rule that decided it), and emails a digest. Claude Haiku 4.5 writes a one-sentence factual summary, and an output check throws out any summary that drifts into advice, price prediction or opinion. A missing summary is acceptable; a slanted one never ships.

Python · FastAPI · Claude Haiku 4.5 · SEC EDGAR · APScheduler

I built the working version for my own prep first. It plans each day to a minute budget, picks problems just above my level in each topic, schedules spaced re-solves and runs a mock interview with a voiced interviewer. When a problem goes badly, Claude diagnoses why, but every rating and the green, amber or red verdict a mentor sees are computed in code, with the reasons attached.

On the multi-user platform I measured cost from the first week. Every model call is logged with its tokens and cost, and live probes showed prompt caching cut the input cost of a 12-turn mock interview by 73%. A teammate set up the staging deploy. The public prototype is a concept demo of that platform, not the working build.

• 73%lower input cost per 12-turn mock interview, with prompt caching

React 19 · Vite · Vercel functions · Supabase · Claude API · ElevenLabs · FastAPI · PostgreSQL · Redis

Every request runs a Claude tool-use loop: Claude decides which tools to call (semantic search over about 1,000 movies, actor and director filmographies, a friend’s taste profile) and how many times, then ranks its picks with a reason for each. It also handles the social part: ask for something you and a friend would both love, or blend your tastes into one shared playlist.

React · Node/Express · Claude tool use · Pinecone · PostgreSQL · TMDB

A machine-learning model gives every transfer bag a live risk score from its connection window, delays, gate or terminal changes and congestion. The dashboard ranks bags by urgency, recommends the fix (prioritize the transfer, expedite customs, escalate to a supervisor), and lets staff log each action from flag to resolution. It retrains on historical outcomes from a CSV upload, so it can adapt to a different airport.

FastAPI · Next.js · TypeScript · ML risk model

PaperTrail finds the likely seed paper, builds its ancestor tree from Semantic Scholar and arXiv citations, and lays it out in learning layers, from the paper itself back to its earliest foundations. Broad prompts like "rag" or "agents" get a short clarifying question first, so the tree starts from the right paper.

React · Vite · D3 · Node/Express · PostgreSQL · Redis · Neo4j

Eight core algorithms, from bubble sort to Dijkstra, each with a step-by-step visualizer, code in JavaScript and Python, and a short quiz. XP, levels, badges and a timed daily quest keep a learner coming back.

React · React Router · CSS Modules