Value map
01.1
AI product strategy
Opportunity mapping, workflow research, value modelling, and technical diligence — a sequenced path from first proof to platform, with the economics attached.
Polymorphic · Applied AISydney · Remote · Est. 2022
Most AI work solves the wrong problem. We find the one worth solving.
Twenty years of shipping production systems, pointed at your operation. We work with your leadership to find where AI actually changes the numbers, build the system around it, and run it live. What you get is a working part of the business, still running after handover.
The plateMap: the latent space of a denoising autoencoder trained on 32 × 32 forms. Its contours are the energy of where real forms land, dented by a slow mass that is felt but never drawn. Pins: heavily damaged forms, placed where the model encodes them, each as tall as its error: the share of the form the model would redraw. Lenses: the same surface read four ways. The model picks the most damaged form it can repair, lifts the map into relief and runs it through its own encode–decode loop, pass by pass: sand cells are restored, outlined cells cleared. A live model, not a recording.
Experience20+ yrs
The senior people who scope your problem are the ones who ship it. No handoff, no juniors.
EngagementEnd to end
Strategy, build, and production ownership in one team — first framing to live system.
First evidence6–8 weeks
A working system on your data, measured against the baseline you run today.
Measured onYour P&L
Margin, cycle time, capacity — success is defined in your numbers, never in model metrics.
01 / Where we start
We start with the business case, then choose the technology.
Most teams arrive with the solution already chosen. We work the other way: map the operation, price the opportunities, and only then decide what to build — a foundation model, a forecast, an optimiser, or nothing at all. Everything after that is aimed at the decision that moves your numbers.
Tool loop
01.2
LLM & agent systems
Grounded assistants, tool-using agents, retrieval systems, eval suites, and the harnesses that keep them observable and recoverable.
Forecast fan
01.3
Applied machine learning
Forecasting, ranking, anomaly detection, optimization, computer vision, and multimodal models built around the decision they improve.
Data layers
01.4
Data & decision systems
Reliable data products, semantic layers, operational analytics, and decision interfaces that turn fragmented signals into action.
Review band
01.5
AI assurance
Risk controls, red teaming, human-in-the-loop design, measurement frameworks, and auditable evidence for high-consequence workflows.
Recovery path
01.6
Technical acceleration
A concentrated strike team for stalled prototypes, brittle systems, architecture choices, and productionisation under a fixed deadline.
02 / Why programmes stall
The problem you were sold is rarely the problem you have.
Vendors solve the problem you asked for. Four failures repeat across every stalled AI programme we are called in to recover — and each one is decided before a line of code is written.
Five steps, and a cheap exit beside each one
Programmes set out along the channel. Most roll off into a pit that looked reasonable in the room; the few that hold every step reach the basin at the end, a system that runs with you.
How to readThe ground is drawn as ridgelines, far to near, each hiding what lies behind it. The path follows a ridge-top channel from your operation to a live system; beside each step the ground falls into a pit named for how programmes stall.
Every step has cheap exits, and each one looks reasonable in the room. Holding all five — the right problem, your real data, the workflow and the estate it has to live in, and someone still accountable after handover — is the whole job.
03 / What you get
Every model needs a system. We deliver both.
A model on its own changes nothing. Around it we build everything that makes it dependable — data, tools, controls, evaluation, and the people who stay in the loop — then run it in production and report against your baseline. You end up owning a monitored, measured part of your operation.
A production system, kept as a ridgeline record
When the inputs drift, the share of requests sent to people rises. The eval monitor catches it, the system falls back and retrains, and the ridges return.
Throughput—Latency p95—Eval pass—Human review—
How to readEach ridge is a slice of time: the model’s confidence across the requests it answered in that window, newest at the front. The part below the review threshold goes to people, in sand. The dashed ghost is the baseline the system was accepted on.
04 / How an engagement runs
Proof in weeks. Scale on evidence.
- 01 / Frame01
Find the leverage
We sit with your leadership, map the operation, and find the decision where improvement is worth the most, whatever the demo looks like.
- 02 / Prove02
Prove on your data
The smallest credible system, built on your data and measured against your current baseline. A clear go / no-go before you commit to scale.
- 03 / Integrate03
Fit your workflow
Data, permissions, interfaces, and human review — designed so your teams adopt the system rather than work around it.
- 04 / Harden04
Run it for real
Observability, evaluation, fallbacks, cost control, and a clear operating model — built to be owned, by your team or ours.
05 / Toolkit
Deep enough to pick the right tool. Disciplined enough to recommend only what fits.
Eleven capabilities we have shipped in production. The methods underneath — Bayesian, causal, deep, or a rules engine — stay in the case unless they earn their place on your problem.
- 01Foundation modelsAssistants and copilots on frontier models — tuned only where it earns its cost
- 02Retrieval systemsSearch over your corpus that returns the right passage, with provenance
- 03Attribution modellingWhat actually moved the number, separated from what merely correlated
- 04Demand & load forecastingWhere volume, price and capacity are heading, with the uncertainty stated
- 05Risk & anomaly detectionThe rare, expensive case caught before it lands
- 06Computer visionInspection, counting and condition monitoring on real imagery and video
- 07Scheduling & allocationRouting, rostering and capacity under the constraints you actually have
- 08Single customer & asset viewEntity resolution and graph structure that give one trusted record across systems
- 09Decision data flowsThe pipelines, semantic layer and contracts a decision system can rely on
- 10Process automationExtraction, classification, routing and exception handling across a live process
- 11Production assuranceEvals, monitoring, rollback and cost control once it is running
From our lab · live at nemradar.com
NEM Radar. The same discipline, running in the open.
Price and spike forecasts for Australia’s National Electricity Market, published every five minutes and checked in the open against the market operator’s own forecast. WEFT, the model behind it, is a transformer whose attention is masked to the wires between regions.
Every NEM price spike since August 2024
Spikes are not spread evenly: they bunch into evening peaks, into winters, and into runs of days, which is what makes them forecastable.
How to readAcross: date, one column per day. Down: time of day, market time. A dot is a spike episode, a run of five-minute prices above $300/MWh in any region, placed where it began; its faint tail is how long it ran. Larger dots reached $1,000 and $5,000. Open squares are prices below zero; the stipple is the day’s ordinary price.
NEM regions5
Forecast cadence5 min
Price horizon2 h
Spike odds30m · 1h · 2h
Spike threshold$300/MWh
Intervals above $300 since Aug 2024 · —
Start here
Bring us the hard, high-leverage problem.
We will tell you what is feasible, what it is worth, and the shortest responsible path between the two — before you commit to anything.