ML-Portfolio
The frontend for ml-portfolio-rho.vercel.app —
a Next.js site presenting around fifty machine-learning tools, each with a
working interface rather than a screenshot. Tabular AutoML, exploratory data
analysis, multimodal RAG, computer vision, document forensics and a set of
security tools.
The models are not here. Every tool calls a FastAPI backend hosted on Hugging
Face Spaces, which is a separate project:
ramleo/ML-Unified.
Accounts & public API. Sign-in is optional (the public tools need no account).
Signed-in users get a profile menu, Account/Settings, and a profile picture; the
owner can issue API keys (/portfolio/keys) that grant access to the public
/v1 REST API and the MCP server — see the site's /docs page.
Layout
| Path | What lives there |
|---|
src/app/tools/ | One directory per tool — 52 of them, each its own route |
src/data/capabilities/ | The tool catalogue, split by domain (pipeline, documents, vision, security) |
src/components/ | Shared UI |
public/thumbs/ | Card photography, generated by scripts/thumbnails.py |
public/demos/ | Recorded walkthrough clips |
e2e/ | Playwright suites — a smoke pass, a per-tool render pass, and a deeper EDA suite |
Running it
npm install
npm run dev
The only variable the site genuinely needs is the backend URL:
# .env.local — git ignores this file
NEXT_PUBLIC_ML_UNIFIED_URL=https://wram1708-ml-unified.hf.space
NEXT_PUBLIC_ML_SQL_URL=https://wram1708-ml-sql.hf.space
Everything else is optional and degrades rather than breaks. Without Supabase
keys the analytics dashboard has nothing to show and /api/track returns 400;
without provider keys (Cohere, Groq, Mistral, Gemini and others) the features
that use an LLM fall back to their non-LLM path. PEXELS_API_KEY is needed
only to source a new card photograph, never to build or run the site.
Tests
npm run lint
npx tsc --noEmit
npm run build # do this first — see below
npm run test:e2e # Playwright, on :3300
Build first, and rebuild after every change. Playwright starts the suite with
next start, which serves whatever is already in .next and never builds it,
and locally it will happily reuse a server you left running. A stale build
therefore fails as a handful of plausible-looking assertion errors rather than
as anything that says "stale build". CI is safe from this — it builds fresh and
does not reuse a server — but a local run is not.
The suites deliberately assert nothing about a backend response. Those Spaces
cold-start and rate-limit, and a suite that goes red for that reason is one
people stop reading.
Two gates run in CI alongside them: scripts/check-file-length.sh fails when a
tracked source file passes 400 lines, and the baseline beside it pins the files
that were already over so they can shrink but never grow.
Licence
Copyright (c) 2026 Ramakrishnasai Wuppalapati. MIT — see LICENSE.
The backend is licensed AGPL-3.0, not MIT: it bundles
Ultralytics-architecture YOLO detectors, whose weights Ultralytics treats as
AGPL derivatives regardless of the runtime serving them. That licence covers
ML-Unified only. This repository is a
separate program that talks to it over HTTP and stays MIT. Both links are in
the site footer, which is where AGPL-3.0 §13 obliges them to be.
public/vendor/paged.min.js is Paged.js v0.4.3, MIT, redistributed with its
licence header intact.
