Private AI for export control and CUI under CMMCThe classification model that is allowed to see your controlled data
Send controlled technical data, or CUI, to a hosted AI API and you can no longer certify that no foreign person saw it. Fine-tune an open-weight model on your own ECCN determinations or CUI marking decisions instead: your data, your weights, your infrastructure.
Keeping data off third-party APIs takes a specific setup. Privacy 3a spells out which features transmit what, and in-house deployment is available.
Private + Owned
Your data, weights & endpoint
1-Line Swap
Drop into the OpenAI SDK
Model-Level MCP
Connect your tools in a click
Your data. Your model. Your jurisdiction.
A hosted API may log, cache, train on, or have staff review what you send it. For a trade-compliance team that is an access-control question you can no longer answer. Keep the data, own the fine-tune, get a private endpoint.
Private & yours, end to end
Your uploads are encrypted and never train a shared model. Flip on delete-after-training and the raw files are erased the moment training finishes. Open-weight fine-tunes are yours: request an export to run anywhere. You own the data, the weights, the training, and the inference.
Drop-in for any dev team
Every model exposes an OpenAI-compatible endpoint. Point the OpenAI SDK's base_url at it. That's the whole integration. Works with LangChain, LlamaIndex, OpenCode, Cursor, or any HTTP client. New teams ship in minutes; existing teams change one line.
Model-level MCP
Give each model its own tools with MCP: a knowledge base, ticketing system, or internal API it can call while it answers. Add credentials once (encrypted), scoped to that model. No servers to run, nothing to redeploy.
The entire integration
from openai import OpenAI
client = OpenAI(
base_url="https://app.infoplatform.ai/api/v1",
api_key="mf_sk_…",
)
client.chat.completions.create(
model="your-model-id",
messages=[{"role": "user", "content": "Hi"}],
)Built for the work that cannot be sent to a third party
A general model has never seen your ECCN determinations, your CUI marking decisions, or how your engineers describe a part. InfoPlatform.ai fine-tunes the open-weight model of your choice on exactly that, so classification reflects your precedent instead of a plausible guess.
Export-Control Classification
The Problem
Misclassification is a violation even in good faith. And the engineering documents you would need to feed an AI are exactly the controlled data you cannot put into a hosted API.
The Solution
Fine-tune an open-weight model on your own determinations and part documentation. It runs on infrastructure you control, the weights are yours, and answers cite the rule they came from.
What this looks like
Drop in a datasheet, get a proposed classification with the controlling text quoted beside it, and correct it once so the next 500 parts inherit your precedent.
CUI Identification under CMMC
The Problem
CUI arrives unmarked or mis-marked, and the contractor carries the burden even though designation is the government's job. And a hosted LLM is not an option: DFARS requires any cloud service touching CUI to be FedRAMP Moderate.
The Solution
The same classifier, pointed at the NARA CUI Registry: an open-weight model fine-tuned on your marking decisions, running inside your boundary. CUI work is an in-house deployment by contract, because for CUI the hosted alternative is not a preference question, it is unlawful.
What this looks like
Drop in a contract or technical document, get a proposed CUI category with the controlling registry text quoted beside it, and correct it once so the next batch inherits your precedent.
Post-Quantum Migration Copilot
The Problem
NIST/NSA deadlines are here and “harvest now, decrypt later” is active, but your cryptographic inventory is too sensitive to send to any SaaS or hosted API.
The Solution
Fine-tune an open-weight model on your own codebase to inventory cryptography (CBOM), risk-rank RSA/ECDH/ECDSA, and generate PQC migration PRs, running entirely inside your walls. You own the weights.
What this looks like
Point it at your repos, get a CycloneDX CBOM with quantum-vulnerable usage risk-ranked, and open migration PRs, without a line of code leaving your network.
Customer Support AI
The Problem
Repetitive tickets drain your support team's time and morale.
The Solution
Fine-tune Qwen 3.5 or Nemotron 3 on your macros and past tickets. Answers like your best agent (your tone, your policies) on a model you own outright.
What this looks like
Paste a real ticket, get a reply in your tone that follows your policies, correct it once, and the correction becomes training data for the next version.
Coding Agents That Know Your Repo
The Problem
Off-the-shelf agents don't know your patterns, libraries, or style.
The Solution
Fine-tune Kimi K2.6 or GPT-OSS on your codebase, then drop it into OpenCode, Cursor, or any harness. Top-tier coding on a model you own outright.
What this looks like
Point your existing harness at a private endpoint with a one-line base_url change, and reviews start citing your conventions instead of generic advice.
Legal Contract Drafting
The Problem
Associates spend hours on boilerplate that follows established patterns.
The Solution
Fine-tune an open-weight model on your clause library, with weights you own and can keep in your own environment. First drafts in your firm's style.
What this looks like
Ask for an NDA against your own clause library and get a first draft in your house style, with non-standard terms flagged rather than quietly accepted.
One Platform, Every Model
Open-weight models are now genuinely competitive on quality. Fine-tune the best of them and own them outright, with no shared-model lock-in.
Kimi K2.6
OpenMoonshot · Open
Elite tool use and long-horizon agent workflows.
Open · you own the weights
DeepSeek V3.1
OpenDeepSeek · MIT
Long-context reasoning & coding specialist.
Open · you own the weights
Qwen 3.5 397B
OpenAlibaba · Apache 2.0
Large-scale open MoE. Multilingual & agentic.
Open · you own the weights
Nemotron 3 Ultra 550B
OpenNVIDIA · Open
Top-tier reasoning at the largest open scale.
Open · you own the weights
GPT-OSS 120B
OpenOpenAI · Apache 2.0
Open-weight reasoning model, yours to run.
Open · you own the weights
Inkling
OpenThinking Machines · Apache 2.0
Open-weight 975B multimodal MoE: text, image & audio. Fine-tune via Tinker.
Open · you own the weights
Every model fine-tunes as the exact weights you pick, no substitute base, from 3B models up to 1T-parameter open MoEs. New open-weight releases are added within days.
From Data to Deployed in 6 Simple Steps
No ML expertise required. Our guided workflow handles the complexity so you can focus on results.
Upload Data
Spreadsheets, docs, or code
Pick a Model
Open-weight, yours to keep
Describe Goal
Tell us what you need
Viability Check
We check if it'll work
Train & Test
Chat, rate, get better
Go Live
OpenCode, Cursor, your app
Before You Train
AI-Powered Feasibility Analysis
Before spending a dollar on GPU time, our feasibility engine analyzes your data quality, task complexity, and expected model performance with a visual radar chart and cost estimation.
- Radar chart scoring across 6 dimensions
- Accurate cost and time estimates before training
- Actionable recommendations to improve data quality
- Go / No-go recommendation with confidence score
Fully Automated
Event-Driven Training Pipeline
When you hit "Train," we spin up a GPU instance, fine-tune your model, encrypt and store the weights, then spin down the infrastructure. Zero idle compute costs.
- Spin up on demand, no standing infrastructure
- Pay per training run, not per hour
- Optionally delete your uploaded data the moment training finishes
Ship & Improve
Test, Deploy, and Improve
When your model is ready to serve, you get a private, OpenAI-compatible endpoint. Test it in the built-in chat, hand your team a one-line base_url swap, connect its tools via MCP, and improve it with your feedback.
- One-line base_url swap into the OpenAI SDK, LangChain, OpenCode, or Cursor
- Give the model your tools with model-level MCP
- Built-in feedback + one-click retrain that learns from your corrections
Everything You Need to Train Custom AI
A complete platform from data ingestion to production inference.
Best-in-Class Open Models
Fine-tune Kimi K2.6, DeepSeek V3.1, Qwen 3.5, Nemotron 3, GPT-OSS, or Inkling. Pick the best open-weight model per project, trained as the exact model you choose.
You Own the Weights
Open-weight fine-tunes are yours. Request an export of your adapter to run in your own cloud, or let us serve them. No shared-model lock-in.
Train Where You Choose
Run training on dedicated GPUs in our own cloud (no third-party training service touches your data) or on Tinker by Thinking Machines Lab. For in-house contracts, the stack runs inside your own infrastructure.
Hosted or In-House, by Contract
Start self-serve on a premium plan, or set up a contract tailored to your business on a call. Either way you own the data and the weights, with a DPA, SLA, and a named engineer on request.
One-Line OpenAI SDK Swap
Every model ships an OpenAI-compatible endpoint. Change base_url in the OpenAI SDK. That's it. Works with LangChain, LlamaIndex, OpenCode, Cursor, or any HTTP client.
Just Upload Your Data
Drag-and-drop spreadsheets, documents, code, or exports. We clean, dedupe, and prepare it automatically.
Viability Check First
We analyze your data quality and task complexity before you pay, so you know it'll work.
Learns From Your Corrections
Rate answers and fix the wrong ones. Your corrections become training data, so a one-click retrain makes the model better exactly where it missed.
Model-Level MCP
Connect each model to your tools with MCP: knowledge bases, ticketing, or internal APIs it can call while it answers. Credentials encrypted, scoped per model, no infra to run.
Transparent Costs
Clear dashboards for training runs, token usage, and inference volume. No per-GPU-hour surprises.
Delete Your Data After Training
Turn on automatic deletion and your uploaded files are permanently erased from our servers the moment training finishes. Your model keeps working; the training data doesn't stick around.
Pricing built for ownership
You are billed per determination: one completed answer, such as a proposed classification for one item. An agent that makes five model calls to reach that answer still counts once, so a more thorough agent never costs you more. Most customers start with a contract shaped on a call. Prefer to self-serve? The plans below train and serve on day one. No free tier, no trial.
Enterprise
Tailored to your business
Determination volumes set by contract
- A contract tailored to your needs and requirements
- Hosted with us, or implemented in-house on your own infrastructure
- Every open model, dedicated capacity, custom volumes
- DPA, SLA, data-residency and security review support
- Named engineer and priority everything
- You own the data, the weights, the training, the inference
Starter
Base platform fee, training included
1,000 determinations included, then $0.75 each
- 2 custom models, fine-tuned on your own data
- 1 configurable agent with human-approval steps
- 3 training runs per month
- Knowledge Base: 50k indexed chunks with cited answers
- Fine-tune the exact model, own the weights
- Email support
Growth
Base platform fee, training included
5,000 determinations included, then $0.50 each
- 5 custom models, fine-tuned on your own data
- 3 configurable agents with human-approval steps
- 12 training runs per month
- Knowledge Base: 250k indexed chunks with cited answers
- Priority support
- Usage dashboards
- Retrain from your feedback
Pro
Base platform fee, training included
20,000 determinations included, then $0.35 each
- 20 custom models, fine-tuned on your own data
- 10 configurable agents with human-approval steps
- 30 training runs per month
- Knowledge Base: 1M indexed chunks with cited answers
- The largest models (100B–1T) unlocked
- Dedicated support
- Fastest processing
Calculate Your ROI
See how much time and money your team could save with a custom AI model.
Estimated monthly savings
$1,400
Assumes 70% of these hours are absorbed by the model. That is an estimate, not a measured result, and export-control and CUI work are scored lowest because every classification still needs a human sign-off.