Scope & eval
We turn "make it better" into a measurable target and build the eval first, calibrated against your experts.
Post-training inside your perimeter.
For organisations where the data can't leave, the GPUs are already reserved and the result has to stand up to audit. POFE runs in your VPC, your data centre or your Databricks workspace — and our research engineers can do the work with you.
Deployment
Three ways to run it. One product.
The same platform in every mode, so a pilot on POFE Cloud moves into your VPC without rebuilding anything.
01
Your VPC
Data plane in your AWS, Google Cloud or Azure account. Training and inference draw down your commitments. The control plane sees run metadata only.
Terraform module · PrivateLink · CMEK
02
Databricks
Native to the lakehouse: Unity Catalog data and permissions, Databricks GPU compute, MLflow lineage and Model Serving endpoints.
UC volumes · MLflow · Model Serving
03
On-prem & air-gapped
Helm chart on your Kubernetes or Slurm cluster, with no outbound connectivity. Offline licence, signed update bundles, your IdP.
Helm · Slurm · offline licence
Governance
Every model has a paper trail.
Risk and compliance teams ask the same questions every time: what data went in, who approved it, how was it tested, what changed since last release. POFE answers them from the lineage graph, not from someone's memory.
- Immutable lineage from source data to endpoint, exportable as a model card
- Approval workflows on datasets, eval sets and deploys
- SSO, SCIM, fine-grained RBAC and project-level isolation
- Audit logs streamed to your SIEM
- Data-residency pinning by region; customer-managed keys
model-card · claims-extractor@dpo-04
basellama-3.3-70b-instruct (Llama community licence)
dataclaims-v7 · 42,880 rows · PII redacted · approved by d.okafor
methodSFT → DPO · LoRA r=64 · 2 epochs
evalsrelease-v3 · F1 0.94 (±0.01) · 0 regressions
judgeκ = 0.86 vs 600 adjuster labels
approvedmodel-risk committee · 2026-09-18
servingeu-central · VPC · CMEK
Managed post-training
White-glove, outcome-priced.
Our research engineers have post-trained models at frontier labs and hyperscalers. In a managed engagement they own the result: you agree the target, they deliver a model that clears it, and your team keeps the pipeline.
Data
Audit what you have, synthesise what you don't, and set up labelling for the edge cases that matter.
Experiments
SFT, DPO and RL across model sizes. We report every run, including the ones that didn't work.
Ship & hand over
Distil, deploy, shadow, then hand over the model, the pipeline and a runbook. We stay on call.
Post-training sprint
Six weeks, one use case, one target metric. The fastest way to prove a tuned open model can replace an API.
Model factory
An embedded team standing up post-training as an internal capability across several use cases, then training your people to run it.
Eval & red-team audit
An independent review of how you evaluate models today: judge calibration, contamination, coverage and release criteria.
Who we work with
Trusted in regulated industries.
3 of 10largest US insurers by premium run POFE in production
2top-5 global banks post-training inside their own VPC
40+healthcare and life-sciences teams training on de-identified data
0bytes of customer training data in our control plane
Enterprise questions
For your architecture review.
Security questionnaires, architecture diagrams and our data processing agreement are available under NDA. Request them here.
Can POFE run fully on-premises or air-gapped?
Yes. The full platform — control plane, data studio, training orchestration, evals and inference — ships as a Helm chart and container images that run on your Kubernetes or Slurm cluster with no outbound connectivity. Licences are validated offline and updates are delivered as signed bundles.
Does POFE work on Databricks?
Yes. POFE deploys into your Databricks workspace, reads training data from Unity Catalog, runs on Databricks GPU compute, logs runs and lineage to MLflow, and can register and serve models through Databricks Model Serving. Governance stays in Unity Catalog.
What does a managed post-training engagement look like?
Week one: we agree the target metrics and build the eval. Weeks two to four: data strategy, synthesis and experiments across methods and model sizes. Weeks four to six: distillation, deployment and handover. You get a model that clears the agreed bar, the pipeline that produced it and a runbook your team can operate.
Who owns the resulting model weights?
You do. Weights, adapters, datasets, eval sets and pipelines produced for you are yours, with no licence fees to keep using them, subject only to the licence of the open base model you chose.
How do you handle sensitive and regulated data?
In VPC and on-prem deployments your data never leaves your environment. We support customer-managed encryption keys, private networking, data-residency pinning by region, PII redaction before training and complete audit trails. Our engineers can work inside your environment under your access controls.
Start with a six-week sprint.
One use case, one target metric, a tuned model in your environment. Tell us what you're replacing and we'll scope it with you.