Every blog automation starts the same way: a schedule fires, you pull in RSS feeds, Google Trends, news and YouTube titles, and then you hand the whole pile to an expensive LLM and ask "pick three good topics".
That step is where the money goes. Most of the candidates are noise — yesterday's weather, a holiday that already passed, a brand announcement with no search intent behind it. You are paying frontier-model prices to say "no" 80% of the time.
So I moved the "no" upstream.
The pipeline
My n8n workflow runs a Korean-language blog with five channels: product comparisons, IT guides, personal finance, trend analysis, and single-product reviews. It looks like this:
schedule → collect seeds (Google Trends / SNS / News RSS / YouTube) → dedupe + normalize → [Jev batch prepare] → [Jev call] → [Jev apply] ← new → prompt build → LLM topic generation → validation → sheet The new part is a three-node detour with a single HTTP Request node inside it. Nothing else in the workflow changed.
What Jev is, in one paragraph
Jev (typesafe/jev-1.14, from TypeSafe AI) is not a chat model. It is a "System One" decision model: you send a state plus a set of questions typed as choice, score, or noul (a yes/no probability), and it returns numbers you can drop straight into an if. Measured in my own testing: 205–539 ms server-side, 0.27–0.92 s round trip. Output tokens are free; you pay per request.
That last property is what makes it usable as a filter: one request can carry up to eight questions and still costs one credit.
Packing eight seeds into one call
The prepare step groups seeds into batches of eight and builds the request body:
const MODEL = 'typesafe/jev-1.14'; const BATCH = 8; // Jev's per-request question limit const MAX_SEEDS = 24; // 24 seeds / 8 = 3 requests = 3 credits per day const state = batch.map(s => `${s.kind} · ${s.title}`); const questions = {}; batch.forEach((s, i) => { questions['s' + (i + 1)] = { type: 'noul', instructions: { candidate_title: s.title, channel_context: CHANNEL_CTX, question: 'Is `candidate_title` a topic worth writing today for one of these channels?' }, criteria: { true: 'concrete, actually being searched today, has information or purchase intent', false: 'vague, trivial, out of season, or no searchable or revenue demand' } }; }); return [{ json: { body: { model: MODEL, state, questions } } }]; The HTTP Request node is deliberately boring:
POST https://thejevai.com/v1/systemone Authorization: Bearer {{ $env.JEV_API_KEY }} Body: {{ JSON.stringify($json.body) }} Two settings matter more than the URL:
- Never Error on,
onError: continue, three retries. A filter that can stop the pipeline is worse than no filter at all. - Seeds are picked round-robin across sources (
trend→sns→news→yt), so one noisy feed cannot eat the whole batch.
Applying the answers
The response comes back keyed by question id, with answers.s1.noul holding a probability between 0 and 1. I keep everything at or above 0.45 and re-attach each answer to its original seed.
Two guards, both learned the hard way:
- A floor on kept seeds. If the filter is too aggressive on a quiet day (
kept < 8), the highest-scoring rejected seeds are restored. The filter narrows the funnel; it must not empty it. - Fail open. If a batch errors out, those seeds pass through untouched. A dead filter degrades to yesterday's behaviour instead of silently producing nothing.
Real numbers from one morning run: 24 seeds checked, 15 kept, 9 dropped, 3 credits, about 900 ms total, 0 failed batches.
The individual judgments looked about right:
| Candidate (translated) | noul | Verdict |
|---|---|---|
| 2026 budget wireless keyboard recommendations | 0.86 | keep |
| Olive Young global sunscreen ranking (K-beauty) | 0.78 | keep |
| Youth savings account guide (Korean gov. scheme) | 0.72 | keep |
| iPhone 18 rumors | 0.63 | keep |
| Samsung 3nm node status | 0.46 | borderline keep |
| Chuseok holiday highway traffic | 0.32 | drop |
| Yesterday's Seoul weather | 0.11 | drop |
The two seasonal rejects are the ones I care about most. A plain embedding-similarity filter would happily keep "Chuseok holiday traffic" because it looks like the travel and finance topics that do well. "Is this worth writing today, in this country" is not a similarity question.
Why not embeddings or a small local model?
Honest answer: I have not measured them head to head yet, and that is part of why I am posting this.
My reasoning so far:
- Embeddings need an anchor and a threshold. Great at dedupe and near-duplicate detection, which I still do separately. Weak at "is this worth writing today", because seasonality and intent are not distances.
- A local classifier needs labelled data I do not have. At 24 seeds a day, I would be labelling them myself for weeks.
- A frontier LLM does all of it well and costs roughly two orders of magnitude more per item at this volume.
The decision model sits in between: no labels, no vectors, a plain-language rubric I can edit without retraining, and a sub-second answer. It also decides whether to spend the big-model call, which is where the saving actually comes from.
The cost shape
Because billing is per request and not per token, batching is the whole game:
- 8 seeds per request x 3 requests a day = 3 credits a day, about 90 a month
- the free signup credit was 10 (expiring in three days), so this needs a paid tier to become a daily habit ($10 for 100k credits in my case)
- dropping
MAX_SEEDSto 16 halves the daily cost
What I would like to hear from you
If you have built this layer before, I would love your take on:
- Threshold tuning. 0.45 came from about 30 samples. Do you drift it per source, or per weekday, or keep one number?
- Batch size. Is 8 the sweet spot for you, or does judgment quality degrade before the limit? I have not run smaller batches against the same seeds yet.
- Alternatives. Is there a local classifier or an embedding-plus-reranker setup that beats a decision model on 20-30 short texts a day, on a cheap box?
- Failure policy. Fail open (mine) or fail closed? I chose open so the blog never misses a day, but that means a broken filter can silently ship bad topics.
I am cross-posting this on Medium and Quora as well, and I will fold the best answers back into the workflow. If you are running something similar, I would rather learn from your scars than collect my own.