Your data. Your model. Your hardware.
We remove patient identifiers before data moves, train private health models in our own lab, and hand you the weights. No third-party AI cloud, ever.
Pt Priya Sharma, MRN 4482913, DOB 14/03/1988, call +91 98201 44312. 34 weeks pregnant, spotting since Tuesday.
Identity removed at the door.
Everything else stays in.
- 01
De-identify at the source
Our de-identification runs inside your app, server or browser and swaps names, numbers and IDs for labels.
- 02
Train and test in our lab
We fine-tune medical models on NVIDIA DGX Spark and test them for safety and accuracy.
- 03
Run it where you choose
Your cloud, your servers, a phone or a browser tab. You own the weights.
Every other option asks you to give something up.
Cloud APIs want your notes. Licensed suites want a per-server subscription. DIY toolkits want your team's time. A private stack asks for none of that.
| Capability | BrandLabs private AIBuilt and run by us | Cloud medical APIsAWS, Azure, Google | Licensed clinical NLPPer-server suites | DIY open sourceAssembled by your team |
|---|---|---|---|---|
| Where it runs | Your hardware, from phone to GPU server | Vendor cloud region | Your servers, licensed per server | Your servers |
| Patient data leaves your network | Never | Sent to the vendor | No | No |
| Cost at 10M notes | No licence fees | Billed per character | Per-server subscription | No licence fees |
| Indian languages | Hindi, Marathi, Tamil, Telugu, Bengali and more | Mostly English | A few, licensed | Mostly English |
| Runs on phones and in the browser | Yes, offline | No | No | Rarely |
| Safety tested before launch | Included | Not included | Not included | Build it yourself |
| Who owns the model | You | The vendor | The vendor | You |
| Who keeps it running | BrandLabs engineers | Vendor support tier | Vendor support | Your team |
Rows describe publicly documented positions as of November 2026. Check each vendor's current terms.
The one thing no cloud API can do: run in a pocket.
The same models run on a laptop CPU, a GPU server, an iPhone or a browser tab. We pick the right one for your users and your data rules.
In your cloud
A REST or gRPC service inside your VPC. Nothing calls out.
On the phone
iOS and Android, on the device. Works offline.
In the browser
Redact in the tab with WebGPU, before anything is submitted.
In our lab
Fine-tune and evaluate on NVIDIA DGX Spark. Ship the weights to you.
Privacy you can audit, not just trust.
All 18 HIPAA Safe Harbor identifiers, Indian phone and ID formats, and a record of every decision.
Signed audit reports
Every run records what was found, which policy applied and what was released.
Context-aware detection
Confidence rises when cues like DOB or MRN sit next to an entity.
Format validation
Checksums and format rules cut false positives on IDs and phone numbers.
Five ways to redact
Mask, remove, hash, shift dates, or swap in realistic fake values.
Zero data movement
No API calls, no licence check-in, no telemetry. Air-gapped if you need it.
Fails closed
Release gates block output when the evidence is missing.
What we don't claim. We don't certify compliance, and we don't ask you to trust published numbers. We rerun every evaluation on your own notes, in your languages, and share the results before anything goes live.
What is BrandLabs private AI?
A private AI stack we build and run for health companies. Identifiers are removed on your own hardware, private models are trained and tested in our lab on NVIDIA DGX Spark, and every release is evaluated for safety and accuracy. It is built from proven open-source projects and our own components, and you own the models.
How is this different from a cloud medical NLP API?
Cloud APIs bill per character and need clinical text sent to the vendor's region, so every new use case becomes a privacy review. A private stack runs in your process, on your servers, a phone or a browser tab, with no calls out and no per-call fees.
Does any patient data reach BrandLabs?
Only de-identified or synthetic data reaches our lab. US patient data stays in your cloud. Identifiers are removed on your side before anything moves, and logs are redacted by default.
Can you fine-tune models on our own vocabulary?
Yes. We fine-tune models on your de-identified data on our NVIDIA DGX Spark machines, evaluate them for safety and accuracy, and hand you the weights. You own the resulting model.
Which languages are supported?
English and major Indian languages, including Hindi, Marathi, Tamil, Telugu and Bengali, with more on request. We measure accuracy on your own notes in each language before launch.
Do we need GPUs?
No. De-identification runs on ordinary CPUs, and faster on Apple Silicon or NVIDIA GPUs. Training and evaluation happen in our lab.
Is it locked in to BrandLabs?
No. The stack is built on permissively licensed open-source software, and you receive the model weights and the deployment. If you stop working with us, everything keeps running.
Does this make us compliant?
No tool does that on its own. We build the controls, measure how well they work on your data, and give you evidence for your lawyers and auditors. We do not certify compliance.
Keep patient data
where it belongs.
Book a 30-minute call. We'll map where identifiers live in your stack and what a private model would take.
contact@brandlabs.app · WhatsApp +91 95947 02015