Four kinds of work that turn the records you keep into something a model can learn from.
Prepare, Protect, Teach, Prove & Run. Each service stands on its own, with a fixed-scope pilot and a published starting price, and each hands its output to the next in a form the next one can use. Most clients start with one and add the second when the first one's numbers justify it.
We do the work between your systems and your model
Most AI vendors sell a model or a platform. The work that decides whether either one succeeds sits in between: getting records out of SharePoint folders, claims systems, ticket queues and scanned archives; making them readable; taking out what must never be learned; turning your experts' judgment into examples; and measuring whether any of it helped.
That is the work we do. We don't sell a model, resell compute, or push a labeling platform. When we run a fine-tune, it is a baseline run to prove the data moved the number, and the weights belong to you.
The twelve services below are grouped into four pillars that follow the order data actually moves in. You can start anywhere, but each pillar assumes the one before it has been handled, by us or by your own team.
How the pieces connect
Records enter at the left as they exist today and leave at the right as versioned, documented datasets with a measured result. Each service picks up at one of the Six Gates and hands off at another.
Inflow Your records: documents, scans, tickets, recordings, tables
Prepare
Protect
Teach
Prove & Run
Outflow Training-grade datasets, evaluation sets, and pipelines you own
The four pillars
Each pillar answers one question. If you can't answer it yet, that's usually where to start.
| Pillar | The question it answers | Services | What you get |
|---|---|---|---|
| Prepare | Do we have data that could teach this task, and can a machine read it? | Readiness Assessment, Document AI & RAG, Curation & Deduplication | An inventory with scores, parsed and indexed documents, deduplicated corpora |
| Protect | What must never be learned, where did each record come from, and what can't we collect? | De-identification, Provenance & Compliance, Synthetic Data | Redacted or pseudonymized data with measured recall, datasheets and lineage, validated synthetic records |
| Teach | What does a correct answer look like, and who in your organization knows? | Fine-Tuning Datasets, Expert Annotation, Agents & RL Environments | Reviewed training examples, rubric scores and preferences, expert trajectories and sandboxes |
| Prove & Run | Did it work, can it run at full volume, and will it keep working next quarter? | Evaluation & Red-Teaming, Data Processing, Managed Data Ops | Golden sets and release gates, pipelines in your cloud, a monthly retraining loop |
All twelve services
Each page lists inputs, outputs, formats, quality gates, turnaround, team and the starting price of the pilot.
Prepare
Find what you have, make it machine-readable, and remove what would teach the wrong lesson.
Protect
Take out what must never be learned, record where every record came from, and generate what you can't collect.
Teach
Turn your experts' judgment into the examples, preferences, and demonstrations that models learn from.
Prove & Run
Measure whether it worked, run the heavy processing, and keep the data improving after launch.
Where clients usually start
Six situations we hear often, and the service that tends to answer them first.
"We have a list of AI ideas and no inventory."
Start with a three-week readiness assessment. You'll know which use cases your data can carry, which can't, and what closing each gap would cost.
"Our assistant gets the tables in scanned PDFs wrong."
Retrieval is only as good as parsing. A Document AI pilot rebuilds one document family with layout-aware parsing, chunking and a retrieval test set.
"Legal won't approve training on anything with names in it."
De-identification with recall measured on a hand-labeled sample gives your privacy team a number to decide on, instead of a vendor's assurance.
"It's in production, and nobody can say if it's getting better."
Build the evaluation suite first: a golden set your experts trust, rubrics, a judge model calibrated against those experts, and a release gate.
"We want a small model that does one job well."
A fine-tuning dataset sprint delivers a task specification, up to 5,000 reviewed examples, a held-out evaluation set and a baseline comparison.
"Our agent works in the demo and fails on real work."
Rebuild one workflow as a sandbox with real tools, seeded data and an automatic verifier, then record how your experts actually do the job.
How to choose where to start
A decision rule we use on scoping calls. Ask the questions in order and stop at the first one you can't answer with confidence.
Can you name the decision or document the model will produce?
If the use case is still "AI for claims" or "AI for maintenance," start with a readiness assessment. It turns a list of ideas into scored use cases tied to real records.
Could you tell a good output from a bad one, in writing?
If your experts would disagree about what "correct" means, no amount of training data will fix it. Start with evaluation: fifty to four hundred cases, a rubric, and a measured baseline.
Can a machine read the records?
Scanned forms, handwritten notes, tables inside PDFs and call audio all need parsing before anything else. If more than a fifth of the source is in that state, start with Document AI or transcription at scale.
Are you allowed to use them?
If records carry patient, customer or employee identifiers, or come from vendors and clients with their own terms, de-identification and a provenance package come before any training run.
Then teach.
With a task, a test, readable records and a clear right to use them, the Teach pillar is where the model actually improves: fine-tuning sets, expert preferences, or agent environments.
Questions about our services
Do we have to buy the services in order?
No. Many clients arrive with the early pillars already handled by their own data team and only need expert preference data or an evaluation suite. We'll check the inputs at the first gate we touch, and if something upstream is missing we'll say so before you spend money downstream.
Do you build or host models?
We don't sell or host models. Fine-tuning sprints include a baseline fine-tune on your infrastructure so we can show the data moved the metric, and the resulting weights are yours. We're glad to work alongside whichever model provider or internal team will run production.
Can you work with the labeling platform or vendor we already use?
Yes. We're tool-neutral and recommend no platform by default. We can work inside your existing annotation tool, or set up an open-source one in your tenant if you don't have one.
What does a pilot include, and what happens after it?
Each pilot has a written scope, a fixed fee, a delivery date and a measurement agreed up front. After it, you can stop, extend the same service to more data, or move to the next pillar. Starting prices for every pilot are on the pricing page.
Where does the work happen?
By default, inside your cloud tenant or on your premises, under your identity provider and logging. Our security page describes the access model, devices, deletion records and what we will never do with your data.
How do ongoing engagements work?
Once a model is in production, Managed Data Operations keeps a standing team sampling real traffic, labeling failures, refreshing evaluation sets and shipping retraining data every month, starting at $8,900 per month.
Keep reading
Bring us one messy data source and one question.
Tell us what the data is and what you want a model to do with it. A data engineer, not a salesperson, will reply within one business day with questions and a proposed first step.