Skillfeed

A tech reading feed ranked to your skills — powered by Jev.

Skillfeed pulls today’s writing from Hacker News, Dev.to, Hashnode, and Lobsters, then scores each piece against a short summary of what you know and care about. You pick the platforms, describe your skills once, and get the best matches first.

Live demo → · GitHub

---

How Skillfeed uses the Jev model

Jev (typesafe-ai/jev) is TypeSafe AI’s System One model: it evaluates a state against typed questions and returns scores and probabilities your app can use directly.

In Skillfeed that looks like this:

1. State — your skill summary (and optional avoid list) plus a batch of article metadata (title, description, tags).
2. Questions — one score question per article: how well does this match the user’s skills? (poor → excellent rubric).
3. Answers — calibrated skill-match scores. Skillfeed sorts the feed by those scores.

We call Jev through the Vercel AI SDK’s experimental_evaluate API on AI Gateway. Articles are scored in batches (UI default: 7) so each request stays under the model’s 32K context window. Multiple questions in one request are evaluated together, which keeps ranking fast and cheap compared with generating long text for every article.

More background: Building a Harness with Jev (LangChain).

---

Features

• Jev skill matching — poor → excellent score rubric against your summary
• Platform multi-select — include only the sources you care about
• Parallel metadata fetch — titles, descriptions, tags (no full-page scrape)
• Batched evaluate — keeps each call under ~32K tokens (UI default: 7 articles)
• SSE progress UI — smart loading with live source counts
• Provider adapter — swap evaluation backends via factory
• Dark-first UI — source icons, ranked list, GitHub link in header

---

Tech stack

Layer	Choice
App	Next.js 16 (App Router)
UI	React 19, Tailwind CSS 4
Ranking	Jev via AI SDK experimental_evaluate + AI Gateway
Validation	Zod
Progress	Server-Sent Events (text/event-stream)
Hosting	Vercel (frontend + API together)

---

Quick start

Prerequisites

• Node.js 20+
• A Vercel AI Gateway API key

1. Clone & install

git clone https://github.com/iikareem/skillfeed.git
cd skillfeed
npm install

2. Environment

cp .env.example .env.local

AI_GATEWAY_API_KEY=your_ai_gateway_api_key_here
# Optional — defaults to vercel-gateway
EVALUATION_PROVIDER=vercel-gateway

EVALUATION_PROVIDER is an app setting (which adapter to use), not a Vercel dashboard field. Today the only value is vercel-gateway.

AI Gateway may require a payment method to unlock free credits. Successful requests only.

3. Run

npm run dev

Open http://localhost:3000.

---

Deploy on Vercel

The UI and backend (/api/rank, /api/sources) are the same Next.js project — one deploy covers both.

1. Import iikareem/skillfeed in Vercel.
2. Set AI_GATEWAY_API_KEY in Project → Settings → Environment Variables.
3. Deploy.

Or from the CLI:

npx vercel
npx vercel env add AI_GATEWAY_API_KEY
npx vercel --prod

Live: skillfeed-xi.vercel.app

---

How ranking works

┌──────────────┐     ┌──────────────┐     ┌─────────────────┐     ┌────────┐
│ User profile │ ──▶ │ Fetch sources│ ──▶ │ Score in batches│ ──▶ │  Sort  │
│ + platforms  │     │ (parallel)   │     │ experimental_   │     │ by score│
└──────────────┘     └──────────────┘     │ evaluate (Jev)  │     └────────┘
                                          └─────────────────┘

1. Collect — up to perSource from each selected platform, deduped, capped at maxArticles.
2. Batch — chunk articles (default UI: 7 per call) for the 32K context window.
3. Score with Jev — typed score questions against state.profile.
4. Sort — highest skill-match first.

Progress streams over SSE so the UI never sits on a blank spinner.

---

API

POST /api/rank

Streams Server-Sent Events. Last event is always complete or error.

Request body

{
  "profile": {
    "summary": "Senior fullstack — TypeScript, Next.js, AI SDK, Postgres.",
    "avoid": "Crypto hype, engagement bait, no-code tutorials"
  },
  "sources": ["hacker-news", "devto"],
  "perSource": 12,
  "maxArticles": 28,
  "batchSize": 7
}

Field	Required	Description
profile.summary	yes	Skills / interests (min 8 chars)
profile.avoid	no	Topics to penalize
sources	no	hacker-news, devto, hashnode, lobsters (default: all)
perSource	no	Articles per platform (default 15)
maxArticles	no	Cap after dedupe (default 40)
batchSize	no	Articles per evaluate call (default 8)

Example

curl -N https://skillfeed-xi.vercel.app/api/rank \
  -H 'Content-Type: application/json' \
  -d '{
    "profile": {
      "summary": "Senior fullstack — TypeScript, Next.js, AI SDK.",
      "avoid": "Crypto hype"
    },
    "sources": ["hacker-news", "devto"],
    "batchSize": 7
  }'

SSE stages: fetching → fetched → scoring → sorting → complete | error

GET /api/sources

Debug metadata fetch (no ranking).

curl 'https://skillfeed-xi.vercel.app/api/sources?limit=10&source=devto'

---

Project structure

src/
├── app/
│   ├── api/rank/              # SSE ranking endpoint
│   ├── api/sources/           # Metadata debug endpoint
│   └── page.tsx               # Skillfeed UI
├── components/
│   └── source-mark.tsx        # Source icons + platform picker
├── sources/                   # One fetcher per platform
├── evaluation/                # Provider interface + Vercel adapter
├── ranking/                   # Collect → score → sort
└── lib/                       # env, http, chunk helpers

• Sources never know about AI — metadata only.
• Evaluation is swappable via EvaluationProvider + factory.
• Ranking owns the pipeline; the HTTP route only encodes SSE.

---

Scripts

npm run dev      # Local development
npm run build    # Production build
npm run start    # Run production server
npm run lint     # ESLint

---

Environment

Variable	Required	Description
AI_GATEWAY_API_KEY	yes	Vercel AI Gateway API key
EVALUATION_PROVIDER	no	Adapter id — default vercel-gateway

Never commit .env.local. Use .env.example.

---

Contributing

See CONTRIBUTING.md for local setup and PR expectations.

---

License

MIT © 2026 iikareem