SortX
SortXTurns messy, duplicated municipal complaints into structured tickets with the department, urgency and location identified, and a person approving each one.
Disha
A phone-first guide that turns a self-taught learner’s skills, goal, device and weekly hours into a short, ordered learning path of free resources, each with a small project.
The problem
A self-directed learner in a Tier-2 city like Bhopal has endless free content but no direction. They can't tell what to learn next, in what order, or whether it leads to real work - content is abundant, but sequencing, feedback, and local relevance are missing. The result: high drop-off, money wasted on generic "roadmap" courses built for other markets, and skills that don't map to work within reach.
How it works
Disha (दिशा, "direction") is a free, phone-first guidance tool. A learner says - in plain language - what they can already do, their goal, their device, and their weekly hours. Claude generates a short, ordered learning path: each step is a real free resource from a curated, licensed catalogue (so links are never invented), paired with a small verifiable project and clear acceptance criteria. It maps the path to real categories of work in Bhopal, Indore, or remote with live job-search links - described honestly, never promised. Learners mark steps done and the path re-plans; the whole plan prints to a one-page summary a mentor can review. It's built on a full teaching engine — retrieval-grounded Q&A over uploaded material and adaptive, auto-generated video lessons - all reasoning powered by Claude. Mobile-first, low-bandwidth, open-source.
Built with
As stated by the team.
Next.js 15 (App Router), React, TypeScript, Tailwind CSS. Reasoning: Claude (Anthropic API, claude-sonnet-4-5) with Zod-validated structured output. Grounding: a curated free-resource catalogue + retrieval-augmented generation (RAG) using on-device embeddings (all-MiniLM-L6-v2 via Transformers.js) fused with BM25, with citations. Teaching video: Mermaid / KaTeX / Shiki visuals + an SVG avatar rendered via headless Chromium (Playwright), muxed with ffmpeg; narration via local macOS TTS. Data: SQLite (better-sqlite3), fully local. Claude is the only paid dependency.
How Claude was used
In the team’s own words, from their submission.
Claude generates a short, ordered learning path: each step is a real free resource from a curated, licensed catalogue (so links are never invented), paired with a small verifiable project and clear acceptance criteria. It's built on a full teaching engine — retrieval-grounded Q&A over uploaded material and adaptive, auto-generated video lessons - all reasoning powered by Claude.
Demo and screenshots
Team
Matmulattention
Individual builders are not listed yet — names are published only with their permission.