The hard part of agents isn’t the model.

Applied data science for agent systems

It’s the dependencies, the context and the people. NotedML builds the data layer underneath agent systems that have to hold up in production, not just in a demo.

See the products Book a working session

PlannerOrchestration TranscriptorAmbient capture DashShared surface RankerPersonalisation MCPeerRegistry

Method

Models are a commodity. The data work around them isn’t.

Every agent project we’ve seen fail, failed on the same three things: nobody modelled the dependencies, nobody measured the output, and nobody could tell you why the system did what it did. We start there.

01 / Model

Map the dependency graph

Before an agent touches anything, we model what depends on what: people, systems, approvals, calendars, records. The graph is the product.

02 / Measure

Evaluate before you ship

Test sets, scored runs and regression gates on every prompt and tool change. If accuracy drops, the build fails, the same way code does.

03 / Monitor

Watch it in production

Traces, drift detection and human review queues wired in from day one. You can always answer why the system reached a given decision.

Products

Five modules. One data layer.

Each runs on its own. Together they share the same dependency graph, the same evaluation harness and the same audit trail.

Planner Orchestration

Calendars that understand consequences.

Scheduling across organisations, not just across inboxes.

Most calendar tools treat a date as a free slot. Planner treats it as a node with upstream approvals and downstream commitments. Move a milestone and it tells you exactly which teams, vendors and people break, and what the nearest feasible window is.

Dependency-aware rescheduling across org boundaries
Constraint solving with human override at every step
Agents and people booked on the same timeline
Change impact shown before the change is committed

Transcriptor Ambient capture

The room, on the record.

Ambient AI for clinical rooms and client sessions.

Transcriptor listens, separates who said what, and turns the conversation into structured notes, coded records and follow-up actions. Nobody types during the session. The clinician or consultant reviews and signs off afterwards, with the audio and the reasoning attached to every line.

Speaker separation and diarisation in noisy rooms
Clinical note structuring against your own templates
Action and commitment extraction for consulting engagements
Audio and transcripts stay inside your tenancy

Dash Shared surface

Dashboards agents can read and write.

Built for a workforce that’s part human, part agent.

A dashboard used to be a place people looked. Now half the readers are agents, and some of them need to write back. Every tile in Dash carries an owner, a lineage and an API, so an agent can query a number, post a revision, and leave a note your team can argue with.

Humans and agents editing the same board, with attribution
Lineage on every metric, back to the source table
Agent-callable read and write endpoints per tile
Comment threads that agents participate in, not just trigger

Ranker Personalisation

Recommendations that know why, not just what.

A hybrid LLM and knowledge graph engine for hyper-personalised ranking.

Ranker builds and caches a knowledge graph over your products or items — attributes, relationships, and the patterns in how people actually choose between them — then uses an LLM to reason over that graph per user, per context. The result reads less like a similarity score and more like a recommendation a good travel agent or shop assistant would make. It’s running today in travel recommenders and ecommerce catalogues.

Knowledge graph built and cached per catalogue, not recomputed per request
Hybrid retrieval: graph traversal plus LLM reasoning over context
Personalisation from session, history and stated intent, not just past clicks
Deployed in travel recommenders and ecommerce ranking today

MCPeer Registry

Every tool your agents can reach, on the record.

An MCP and skill registry with governed access and promotion policy — built for product managers, not engineers.

Agents get their power from the tools you connect to them, and in most organisations nobody can list those tools, say who approved them, or name the version in production. MCPeer is the registry that answers all three. Servers and skills are published once, versioned, and promoted through environments under policy. Access is granted to a team or an agent, reviewed, and revoked — from a console, by the person who owns the product.

Every MCP server and agent skill registered once, with a named owner
Version pinning and dev → staging → production promotion gates
Access granted per team and per agent, then reviewed and revoked
An audit trail of who connected what, when, and on whose approval

Why a registry is a governance control →

Founder

Rob Cooper

founder, NotedML

“Agentic systems need rigour and intuition in equal measure. The real work is putting the right AI tooling and governance in the hands of the people who own the product, not locking it away with engineering teams.”

Rob has spent the last four years building multi-agent systems for organisations that can’t afford them to misbehave, across travel, retail, sport and resources.

His work has taken AI agents from demo to production for some of Australia’s largest brands, delivered alongside major cloud partners and recognised in national AI awards. The same problems kept showing up project after project: nobody modelled the dependencies, nobody measured the output, nobody could explain why the system did what it did.

NotedML is what came out of that pattern. He built the modules to solve it properly, once.

Brisbane, Australia

Writing

What we’ve had to work out the hard way.

Frameworks, primers and field notes on getting agent systems past the demo. Written for the people who have to sign off on them.

Primer / Aug 2026

Contact

Tell us what you’re trying to coordinate. We’ll tell you if it’s a data problem or a model problem.

Email hello@notedml.com