Calibrated probabilities
Not a thumbs-up classifier. Real softmax values from a fine-tuned encoder model — pick the threshold that matches your appetite for false positives.
spam classification · v1
Stop spam
at the API.
Updated May 12, 2026
Drop in one endpoint. Get a probability back.
No queues. No models to host. No surprises.
10,000 requests/month free · no credit card · two-line install
curl -sX POST https://api.siftfy.io/v1/predict \
-H "X-API-Key: $SIFTFY_KEY" \
-H "Content-Type: application/json" \
-d '{"text": "FREE iPhone! click http://bit.ly/x to claim NOW"}'
# {
# "spam_probability": 0.987,
# "likelihood": "high"
# }99.4%
accuracy on public test set
<10ms
p99 latency, CPU only
1
endpoint to learn
270MB
model · DistilBERT-class
How it works
01
Create a key
Sign in with your email — we send a one-time link. Generate an API key from the dashboard.
02
Send a request
POST any text to /v1/predict. The response is a softmax probability between 0 and 1.
03
Threshold & ship
Pick your threshold (we suggest 0.85). Block, queue, or quarantine — that's up to you.
Built for an inbox that doesn't wait.
Single round trip
No queues, no async webhooks, no streaming. One HTTPS request, single-digit milliseconds, response inline.
Per-key everything
Every key has its own usage counter, last-used timestamp, and revocation. Issue keys to specific apps; revoke any one of them in a click.
Deterministic by design
Model version is pinned per release. Same input, same probability — no silent drift on your spam filter when we retrain.
In production at
Frequently asked questions
- How accurate is the model?
- We benchmark at 99.4% accuracy on a public spam test set. Because the response is a calibrated softmax probability rather than a hard label, you can pick a threshold (we suggest 0.85) that matches your appetite for false positives.
- What language(s) does it support?
- The active model is trained primarily on English. Non-English inputs still classify, but accuracy degrades. If you need a specific language we can swap the underlying model — email us.
- How fast is /v1/predict?
- Single-digit milliseconds for SMS-length text on CPU. Most calls round-trip in under 10ms p99 from the same region. There are no async queues, no webhooks — one HTTP request, one response.
- Do you store the text I send?
- No. We log request metadata (timestamp, status, duration, tenant id) but the request body is not persisted. If you need contractual confirmation in writing, get in touch.
- What happens if I exceed my quota?
- On the free tier you get 10,000 requests/month and 60/minute. Cross either and the API returns 429 with a Retry-After hint. Pro lifts the cap to 1M/month included plus $0.0001/req metered.
- Can I self-host?
- We run a managed service. For very large volumes or air-gapped requirements we can talk about a dedicated deployment — email hi@siftfy.io.
Drop it in where you already are.
Eight integrations, each with a real runnable example. Pick the one that matches your stack — or read the /v1/predict reference and roll your own in 15 lines.
- Contact formsServer-side classification per submission. No CAPTCHA, no honeypot.
- CommentsScore every new comment on a calibrated 0-1 axis before it goes live.
- SignupsCatch fake-account registrations on the natural-language fields.
- CAPTCHA alternativesReplace reCAPTCHA friction with one server-side call. Same protection, no UX tax.
- WebflowFront Webflow's form-submission webhook with a 30-line worker.
- Ghost CMSFilter spam member signups via Ghost's webhook + Admin API.
- Static sitesAstro, Next.js, Gatsby, SvelteKit — one serverless function does it.
- Headless CMSSanity, Contentful, Strapi — generic webhook pattern, per-type field mapping.
Two lines from now,
you're filtering spam.
10,000 free requests every month. Pro is $9 + $0.0001 per request after that.