Frontier Accuracy.

Special.Small.Special. Owned.

Purpose-built models that run inside your infrastructure.

95.6% accuracy on held-out tool calls · 88.1% F1 on Privacy-Bench documents · Open-source

Benchmarked, not promised

18.8×

faster than a generative LLM judge

75.4% exact accuracy vs. human labels, scoring a response in one pass instead of a long generation.

RubriX 8B

5,000+

Downloads per month

50+

Models

10+

Model families

10+

Production AI systems deployed

The trust layer for enterprise AI.

Three jobs every AI rollout needs — or browse by industry.

Personal data, credentials, prompt attacks and offensive content, stopped before they leave your perimeter.

PII Extractor

Every personal detail, located precisely.

68M · 32M · 17M · 4M

PII Extractor 68M88.1%
PII-Tracer (596M)86.9%
Presidio (spaCy lg)82.4%
GLiNER2-PII (205M)80.8%

Exact-match span F1, Privacy-Bench docs (656 human-annotated pages)

0.06 false values per 1,000 characters on text with no PII (68M), against 1.6–3.6 for the other models.

Indic PII Extractor

Personal data, found across 12 Indian languages.

58M · 36M · 17M

Indic PII Extractor 58M98.6%
Chitragupta (with its rules)94.1%
GLiNER-PII large86.7%
IndicBERT-PII72.7%

Typed-entity F1, Ylemis India-PII (10,000 texts)

19.9% of PII-free documents wrongly flagged on IndicPIIDetX (8,028 docs), against 53.4% for Chitragupta's rule-based baseline and 61.0% for IndicBERT-PII.

Secrets Detector

Every secret, found.

32M · 17M

Secrets Detector 32M0.861
Gitleaks 8.30.10.678
detect-secrets 1.5.00.29
TruffleHog 3.99.20.228

Span F1, real-world issue reports (949 secrets)

Prompt Injection Detector

Attacks on the model, caught before they run.

68M · 32M · 17M

Prompt Injection Detector 68M0.987
PIGuard 184M0.351
deepset 184M0.267
Prompt Guard 2 86M0.254

Average precision, CBX Benchmark (2,506 matched attack/clean pairs and hard negatives)

0.861 average precision on real injected web pages (BrowseSafe-bench, 3,680 rows), and 93.7% of attacks caught at 2.8% false alarms on the in-house v2.2 test set (5,584 rows).

Indic Profanity Detector

Offensive content, caught across 12 Indian languages.

58M · 36M · 17M

Indic Profanity Detector 58M0.89
TF-IDF + logistic regression0.87
Lexicon rule (598 terms)0.7

Offensive-content F1, gold test set (3,980 texts, two-LLM labels)

3.5% of clean text wrongly flagged, against 7.6% for TF-IDF. The lexicon rule flags less clean text (2.9%) but scores only 0.36 F1 on hate speech and has no insult label at all.

Why ownership pays

Your edge stays with you

Every model runs on your infrastructure, so your data improves your systems alone.

An asset, not a subscription

Open-source weights you keep, tune and build on, at the same cost however much you use them.

Runs where your data lives

On CPU or GPU, in your cloud or on-premise, beside the systems you already run.

Where it runs

Pre-commit hook

Run the secrets detector on every diff before it leaves your machine.

Log pipeline

Run the PII extractor over logs and tickets before they hit storage.

Agent eval harness

Run the tool judge and RubriX in CI, on every agent change.

Deployed and tuned by the team that built them.

Deploy

On your cloud or on-premise, wired into your pipelines and CI.

Fine-tune

Adapted to your documents, labels and rubrics.

Build the next one

A purpose-built model for the job your business needs next.