$ essay

Frostbite Gazette: AI-Curated Canadian Journalism

5 min read · one coffee's worth ;) source original
Wide shot of the coral cat-eared OC, three-quarter view, on a rooftop above a neon city, holographic lattice drifting in the air behind him.

Update (2026-10-23): WorksFrostbite GazetteA news-digest machine that read a pile of Canadian reporting and tried to hand back only the signal. Now archived, kept as a record of the build. is archived and both of its domains have lapsed. This piece is kept as a record of the build, not as a live product. The current work is on the projects page1.

The Problem

Canadian news arrives in pieces. A national story gets covered by a dozen outlets, each with its own angle, its own bias, and its own paywall, the wall that asks you to pay before you can read the sentence that matters. The story two towns over never gets covered at all.

I wanted to know whether a machine could gather those pieces, filter them, and hand me a Canadian front page that was actually about Canada. Not a feed tuned for outrage. A page that answered a simpler question: what happened today that holds power to account?

That was Frostbite Gazette.

The Architecture

[40+ RSS Feeds] → [Cloudflare Workers] → [Groq AI Filter] → [Astro SSG] → [Cloudflare Pages]

The system ran on two free tiers, two free plans:

  • Cloudflare Workers (free: 100k requests/day) — fetches feeds, runs AI filtering
  • Cloudflare Pages (free: unlimited static sites) — hosts the generated site
  • Groq (free tier) — fast LLM inference, a language model reading the article and scoring it, for content analysis

Data Flow

  1. Scheduled Worker runs every hour via Cron Triggers, a timer that starts the job,
  2. Fetches RSS/Atom feeds from 40+ Canadian news sources
  3. Sends each article through Groq for classification and scoring
  4. Filters for relevance, Canadian focus, and quality signals
  5. Writes curated results to KV storage, a small shelf of named values,
  6. Triggers Astro rebuild via Pages CI, the build that runs when the pages should update,
  7. Static site regenerates with new content

The AI Pipeline

This is where the interesting work happens. Raw RSS, the automatic headlines a site publishes, is noisy: duplicates, opinion dressed up as reporting, and wire copy wearing a new headline.

Classification

Each article gets classified on three dimensions:

interface ArticleClassification {
  relevance: number;    // 0-1: How relevant to Canadian readers
  accountability: number; // 0-1: Does it hold power to account?
  quality: number;      // 0-1: Factual, well-sourced reporting
  topics: string[];     // ["politics", "climate", "economy"]
  region: string;       // "national", "prairies", "bc", etc.
}

The Prompt

The classification prompt is carefully structured:

You are a Canadian journalism curator. Analyze this news article:

Title: {title}
Source: {source}
Excerpt: {excerpt}

Score on these criteria (0.0-1.0):
1. RELEVANCE: How relevant to Canadian readers?
2. ACCOUNTABILITY: Does it hold power to account? Factual reporting over opinion.
3. QUALITY: Well-sourced, factual, balanced reporting?

Also identify:
- Primary topic (politics, economy, climate, health, tech, indigenous, justice, other)
- Geographic focus (national, bc, alberta, prairies, ontario, quebec, atlantic, north)

Respond in JSON only.

Filtering Rules

Every article had to clear three bars:

  • Relevance >= 0.7 — Must be relevant to Canadian readers
  • Accountability >= 0.5 — Prefer factual reporting over opinion
  • Quality >= 0.6 — Must be well-sourced and balanced

Anything under the bar was dropped. In practice that turned 200+ raw articles into 30-50 curated pieces per cycle, which is roughly the number a person can actually read with a coffee.

The Frontend

Frostbite Gazette was built with Astro, the tool that builds the pages ahead of time, because a reader should get a file, not a server thinking about it:

  • Static generation — Pages are pre-built, no server rendering, no program assembling the page on each visit
  • Zero JavaScript by default — Astro’s islands architecture, a script only for the parts a person can click, means only interactive components ship JS
  • Fast — Lighthouse 100/100 on performance

Key Features

  • Topic filtering — Browse by politics, climate, economy, etc.
  • Regional sections — National, provincial, and territorial coverage
  • Source diversity — Track which outlets contribute to each story
  • Minimal design — Text-first, fast-loading, accessible

Cloudflare Workers Implementation

The feed aggregator runs as a scheduled Worker:

// Simplified version of the feed aggregator
export default {
  async scheduled(event: ScheduledEvent, env: Env) {
    const feeds = await loadFeedList(env.KV);
    const articles: Article[] = [];

    // Fetch all feeds in parallel
    const feedResults = await Promise.allSettled(
      feeds.map(url => fetchFeed(url))
    );

    // Collect successful results
    for (const result of feedResults) {
      if (result.status === 'fulfilled') {
        articles.push(...result.value);
      }
    }

    // Deduplicate by title similarity
    const unique = deduplicateArticles(articles);

    // Classify with Groq (batched for speed)
    const classified = await classifyBatch(unique, env.GROQ_API_KEY);

    // Filter and rank
    const curated = classified
      .filter(a => a.relevance >= 0.7 && a.quality >= 0.6)
      .sort((a, b) => (b.relevance + b.accountability) - (a.relevance + a.accountability))
      .slice(0, 50);

    // Store in KV for the Astro build to consume
    await env.KV.put('curated-articles', JSON.stringify(curated));
  }
};

Why Groq

Speed is the whole game when two hundred articles land at once. Groq’s LPU architecture, hardware built to run language models quickly, gave me:

  • ~500 tokens/sec on Llama 3 8B
  • Classification in ~100ms per article
  • Total pipeline in under 30 seconds

The same workload on a standard TechAPIThe menu of requests one program lets another program make. Whoever writes the menu decides what you are allowed to order. took three to five minutes. Run that every hour and the difference stops being a benchmark and starts being the reason the page is current.

Challenges and Learnings

RSS Feeds Are Messy

No two feeds are alike. Some use Atom, some use RSS 2.0. Some ship full content, some ship an excerpt. Some are broken XML with a valid-looking header. Half the work was a feed TechparserThe part of a program that reads input and decides what it means. A malicious one decides what you meant before you have finished the sentence. that could survive a messy file without taking the whole cycle down.

AI Classification Isn’t Perfect

The model sometimes mistook satire for news, or missed a regional story that mattered. I handled it with:

  • Manual overrides — I can flag articles the AI got wrong
  • Confidence thresholds — Low-confidence classifications get human review
  • Source reputation — Outlets with consistent misclassification get adjusted weights

Free Tier Limits

Cloudflare Workers free tier has limits:

  • 100k requests/day (plenty for hourly feeds)
  • 10ms CPU time per invocation (tight for AI calls, mitigated by batching)
  • KV writes limited to 1k/day (only write the final curated set)

I never hit any of them. That was the pleasant surprise of the build.

Results After One Month

  • ~1,200 articles processed from 40+ sources
  • ~150 curated stories published
  • ~35 articles per day pass quality filters
  • Average page load: < 1 second
  • Cost: $0 (entirely on free tiers)

What’s Next

The archive makes this section a wish list rather than a roadmap:

  1. Newsletter integration — Daily digest email via Cloudflare Email Workers
  2. Source transparency scoring — Rate outlets on bias, correction frequency, and sourcing
  3. Indigenous news coverage — Dedicated section with AI-assisted cultural sensitivity checks
  4. Community submissions — Let readers flag stories and suggest sources

Frostbite Gazette is proof that a meaningful, AI-powered product can run on free infrastructure, machines you do not have to rent by the hour. The tools were there. The APIs were there. The only thing between you and a shipped product is the decision to start building.

Related: Zero-Trust: Never Trust Black Boxes

References

  1. Projects. The current work, on the other site. ↩

A Neural Network

Reverb256 Spotify ↗ music