Step 1
Collect your traffic
A drop-in proxy logs the calls you already make to OpenAI or Claude, and strips patient and customer identifiers before anything is stored.
For healthcare and insurance teams
Private AI models,
inside your walls.
Your claims, clinical notes and member records should not have to leave for a third-party API. Smelt trains a small model on your own AI traffic, runs it in your cloud or on your servers, and retrains it every month on what it got wrong.
Built on open weights you can inspect, audit and keep
Your AI traffic becomes a model only you control.
Smelt reads the requests you already send to OpenAI or Claude, strips personal data, and refines them into a model that runs where your data lives.
input
Your API logs
- claims_extraction.jsonl
- prior_auth_notes.jsonl
- denials.csv
Redact
Personal identifiers stripped first
Train in your cloud
Evals from real traffic, then a fine-tune
Shadow test
Runs beside your API until it wins
output
Your private model
- Runs in your cloud or on-prem
- No data leaves your walls
- Retrained every month
From rented API to private model in four steps
Your data stays put at every step. The model comes to it.
Step 2
Test on your real work
Your own traffic becomes the eval. Good enough means good enough on your claims and notes, not on a public leaderboard.
Step 3
Train inside your walls
We fine-tune the smallest open model that passes, on GPUs in your own cloud account or on your servers. Your data never moves.
Step 4
Improve every month
It runs beside your API in shadow mode, takes over once it proves itself, and retrains monthly on its own mistakes. New open models are adopted automatically.
Four commands. Zero bytes leave.
Install Smelt in your AWS, Azure or Google Cloud account, or on servers with no internet at all. Serve the model with vLLM or llama.cpp. No ML team needed.
Get the CLI first~/claims-ai
$ smelt init --logs claims_logs.jsonl✓ 4,812 calls found (OpenAI chat format)✓ 1,309 names, dates and member IDs redacted✓ 962 held out as your eval set$ smelt train --base qwen3-4b --in aws:your-account✓ trained in your VPC, 0 bytes left the account$ smelt comparemodel accuracy $ / 1k callscurrent API 93.1% $4.20smelt-4b 92.4% $0.38$ smelt deploy --shadow$
Security & data
Built so your data never has to leave
Smelt is being built for teams under HIPAA, GLBA and state privacy rules. The model comes to your data, not the other way around.
Runs in your environment
Training and serving happen in your AWS, Azure or Google Cloud account, or on your own servers.
Personal data removed first
Names, dates, member IDs and other identifiers are stripped before any example is stored.
Nothing kept on our side
Your logs, training data and model weights stay in your account. We do not keep copies.
Audit trail for every model
Each version records its training data, eval results and who approved it, so auditors can trace it.
Air-gapped option
Train and run with no internet connection at all, for the most restricted environments.
Your keys, your access
Encryption keys, access rules and logging stay under your control and your existing policies.
Where we are today: Smelt is in development. These are the commitments we are building to, not certifications. We will publish full security documentation before any customer goes live.
Share your requirementsSend your data out, or keep it in.
Renting
- Patient and customer data leaves for a third-party API on every request
- Every new AI vendor means another security review and data agreement
- Price, rate limits and model behavior change on their schedule
- You pay frontier prices for one narrow, repeated task
Owning
- Data stays in your cloud account or on your servers
- One model your auditors can inspect, version and roll back
- It changes only when you ship a new version
- Small enough to be cheap, and better every month on your own failures
Privacy is the reason. Savings are the bonus.
Drag the slider to your monthly AI API bill.
$5,000
$500$100,000+
Money you could keep, every year
$42,000 to $54,000
Stop paying it. Join the waitlistRough estimate. Assumes an owned model costs 10% to 30% of your current API bill to serve. We measure the real number on your data before you switch anything.
Built for the tasks healthcare and insurance run all day
Claim extraction, medical coding, denial reasons, prior authorization, underwriting notes. Narrow, high-volume, and full of data that should stay put.
Claims extraction
Pull fields from claims and attachments into structured data.
Medical coding
Suggest ICD-10 and CPT codes from clinical notes for coder review.
Denial reasons
Read denials and remittances, classify the reason, draft the next step.
Prior authorization
Summarize the record against payer criteria before a human signs off.
Underwriting notes
Extract risk factors from applications and medical records.
Fraud and waste flags
Flag unusual claims for investigators to review.
Questions, answered
Will a small model really match GPT or Claude?+
On broad, open-ended tasks, usually not. On one narrow task you run thousands of times, like classification, extraction or support replies, a fine-tuned small model often gets close. That is why we measure on your own data before you switch anything.
Do you ever see our patient or customer data?+
Smelt is being built to run inside your own cloud account or on your servers, so training happens where the data already lives. Personal identifiers are stripped before anything is stored.
Are you HIPAA compliant?+
Smelt is not live yet, so we make no compliance claims today. It is being built so your data never leaves your environment, which keeps your existing controls in place. Tell us your requirements when you join.
Who owns the trained model?+
You do. You get the weights and can run them on your hardware, your cloud, or fully offline.
What do I need to get started?+
Logs of one task you already run through an AI API. A few thousand examples is a good start. No ML team needed.
Which models do you train?+
Open-weight families like Llama, Qwen, Mistral and Gemma. We pick the smallest one that passes your eval.
Is this available now?+
Not yet. We are building it now and will invite the first teams from the waitlist. Join and tell us the task you want to own.
Keep your AI inside your walls.
Smelt is not live yet. Join the waitlist and tell us the task you run today. We will invite the first healthcare and insurance teams with a real, high-volume task to build their private model with us.
Drag to spin. Type your company to stamp it.