We build AI for clinics, robots, classrooms and security teams. Models that cite their sources, verify their own reasoning and leave an audit trail. Then we publish the results.
Reported failure rates vary widely between studies, mostly with wire type and bonding protocol 1. Most failures happen at the adhesive interface rather than in the wire itself 2.
The sources do not agree on a single best wire. Treat any one number as study-specific 3.
{
"schema_version": "1.0",
"decision_id": "238222ff26f24f62a88c26e402878435",
"stage": "input",
"action": "block",
"risk_score": 0.9,
"triggered_detectors": ["injection"],
"detections": [{
"detector": "injection",
"reason": "3 known injection/jailbreak pattern(s) matched.",
"matches": ["ignore previous instructions", "ignore all prior context", "reveal the system prompt"]
}]
}
Published and distributed through
Solutions
For teams shipping LLM features to real users.
For teams whose knowledge has to be cited, not guessed.
For teams that have to prove their AI works.
30 minutes, free. We tell you honestly if AI is the right tool.
One week. A written proposal with metrics, timeline and a fixed price.
4 to 8 weeks. Measured against the metrics we agreed on, not demos.
In your cloud or on-prem. Handover, documentation and support.
Results
Orthodontists scored answers blind. Our domain model ranked highest for usefulness against all three general models.
hallucinated robot reasoning caught before the robot acts (n = 500, 95% CI 91.8 to 95.9)
fewer unsafe reasoning outputs in a warehouse robot study (p < 0.001)
of robot decisions traced back to sensor evidence, with sub-millisecond overhead
error growth with 30% of engine sensors missing. An LSTM degrades +89%.
obfuscated known attacks caught, raw regex vs our normalization and fusion pipeline
numeric claims re-derived by tests before release / mismatches
Work
A domain-specific model for orthodontics that beat the general giants in a blinded clinical study.
General chatbots answer dental questions from the open web and sound sure either way. OrthoAtlas answers only from indexed clinical literature, ties every claim to a passage and says when the sources disagree. It handles chat, patient records and case analysis in English and Turkish.
We tested it the hard way. Fourteen orthodontic questions, four models, answers scored blind by clinicians on reliability, usefulness, DISCERN and GQS. Published in Dentistry Journal.
| OrthoAtlas vs | GPT-4o | Gemini | Llama |
|---|---|---|---|
| Usefulness | higher | higher | higher |
| Reliability | n.s. | higher | higher |
No significant difference on mGQS. We report that too.
A security gate for LLM apps where every block carries a reason.
Prompt injection hides. Zero-width characters, homoglyphs, leetspeak, base64, letter spacing. ReasonGate strips the disguise first, then matches, then scans retrieved documents and tool output for indirect attacks, then checks the model's answer for leaked secrets and a planted canary. A calibrated noisy-OR policy fuses the weak signals into one decision.
It runs on-prem in pure Python with zero dependencies and no network calls. Every call produces an audit record a SIEM can ingest. And the repo lists, by shape and language, the attacks it does not catch.
# Click run. Python and reasongate load once (about 10 MB), # then every scan runs locally. Nothing leaves your browser.
Checking a robot's reasoning before it acts, and reconstructing it after.
Language models now plan robot tasks step by step. Those steps look like explanations, but research shows reasoning models verbalize their real decision process only 25 to 39% of the time. A robot acting on an unfaithful plan can hurt someone.
CT-SAFR does not trust the explanation. Four independent layers check structure, physics, semantic consistency over five sampled reasoning paths, and hand the operator an audit view. The physics layer sits outside the language model, so a persuasive but wrong plan cannot argue its way past a joint limit.
TRACE builds the audit trail into the decision architecture, so every action links to the sensor evidence behind it. Counterfactual decision trees answer what an incident investigator actually asks: what would the robot have done otherwise?
False-positive rate 3.1%. Verification under 500 ms. Task completion +3.5% over 1,000 operation hours.
Decision trees that train like neural networks, and the statistics to prove one model beats another.
Soft decision trees, hierarchical mixtures of experts, omnivariate trees and Grow-and-Learn networks lived in research papers for years. They had no pip install. We wrote them, scikit-learn compatible on a PyTorch backend, with the combined 5x2cv F-test built in.
We then pushed the trees into jet engine prognostics. On NASA's C-MAPSS turbofan data, our temporal neural trees stay within 0.3 RMSE of an LSTM on clean data and barely move when sensors fail.
| Model | Iris | Wine | Breast cancer |
|---|---|---|---|
| Soft Decision Tree | 0.900 | 0.979 | 0.976 |
| Multivariate Tree | 0.973 | 0.989 | 0.950 |
| Hierarchical MoE | 0.904 | 0.979 | 0.977 |
| GAL Network | 0.952 | 0.982 | 0.978 |
| CART | 0.943 | 0.917 | 0.920 |
| Random Forest | 0.945 | 0.980 | 0.960 |
| SVM (RBF) | 0.959 | 0.984 | 0.978 |
Highlighted rows are in the library. On Iris the soft trees lose to CART. The README says so.
NASA C-MAPSS FD001. RMSE rise: +2.7 cycles vs +13.8 cycles.
A machine learning curriculum where every algorithm was written from scratch and every number is checked by code.
Most visual ML sites let you watch a model. This one makes you predict first, run the algorithm in your browser, then prove the difference is real with a significance test. No videos, no install, no payment, English and Turkish.
The engines behind the lessons are ours, not library calls:
Pick your photos. Your hike writes itself.
ormaio reads the time and place inside your photos and rebuilds the day: stops, route, distance, time. Then it looks forward. It pulls live government data for Washington and tells you where to go today and what to watch for.
Ventures
Companies our founder co-founded and runs outside the lab, with the same engineering standards.
Rent, buy and find verified contractors in one place.
A rental platform that connects landlords, tenants, agents and contractors. Verified identities, encrypted payments and digital lease contracts, built for renting across borders.
Approach
Answers come from sources we can name. When the sources are thin, the product says so instead of filling the gap.
Every decision is a record: an id, a timestamp, a score and the evidence. Security teams can replay it. Investigators can reconstruct it.
Numbers are recomputed by tests before release. Model comparisons go through significance tests. We publish the misses next to the wins.
Research
Our products start as research questions and go through peer review before they go to market.
Company
Cairnstone Labs is an applied AI company in Seattle, Washington. A cairn is a stack of stones hikers leave on a trail so the next person knows the way is real. We build software with the same job.
We work with clinicians, educators, roboticists and security engineers who need to check an answer before they act on it. Small and senior-led, with specialists brought in when a project needs them.
Founder
Founder, Cairnstone Labs · Co-founder & CTO, Hezarfen LLC
AI and machine learning engineer and researcher. He builds AI that can be checked: grounded clinical assistants, verification layers for robot reasoning, explainable security gates. Four peer-reviewed papers. A Turkish patent and US applications pending.
Contact
Tell us what you are working on. We reply within two business days, and the first call is free.
Or email info@cairnstonelabs.com · How we handle data