Pentatonic LabsApplied AI researchLondon and San Francisco
Run. Remember. Learn. See. Act.
Pentatonic Labs is an applied AI research lab that builds and deploys AI that can run, remember, learn, see and act, on hardware inside your business.
Continue in Physical AI for Commerce →Physical AI for Commerce
- Footwear
- Beauty
- Toys
- Drinkware
- Apparel
Seesa
- Startups to FTSE 100 companies
Working with
The idea
The next leap in AI comes from where models run and how they learn. We take AI to where the data is and the work happens: the office, the warehouse and the store.
Applied frontiers
One record. Two applied frontiers.
Physical AI for Commerce
Models read products from the design file to the factory to the returns counter.
OutcomeBetter products. More value recovered.
Enter Physical AI for Commerce →Seesa
A company's own AI that lives in its building, remembers how the company works, and acts in ways people can check and undo.
OutcomeLess chasing. Faster decisions.
Enter Seesa →The difference
Today's options, and ours.
- Today: cloud copilots
Rent you their intelligence.
It stays in their cloud.
- Today: on-prem model stacks
Install a static model.
It stays as installed.
- Pentatonic Labs
Builds and deploys a model that becomes yours.
Running in your building today. Learning from your own record is in build.
Programmes
Five programmes. Two disciplines underneath.
Shipped and running in customers' buildings: Run, Remember and Act. Live on our reference kiosk: See. In build: Learn, our compiled models.
ShippedLive pilotMeasuredIn buildExploring
- Run
Resident Inference
Strong open-weight models, served on desk-side and edge hardware inside the business.
- Resident AttentionShipped for conversations
- Foreground PriorityShipped
- Exact DraftingShipped
- Zero-Token RoutingLive pilot on our own box
- Remember
Grounded Memory
A record in which every fact traces to its source and stays true over time.
- Lifecycle LedgerShipped
- Schema-Locked ExtractionShipped
- Self-Healing MemoryIn build
- Rank-First RetrievalMeasured
- Learn
Compiled Models
Models that become the customer's own: rebuilt from their record, in their building, for them alone.
- Memory CompilationIn build
- Parity GateIn build
- Paired GradingIn build
- Hybrid Data PipelineIn build
- See
Lifecycle Vision
Product intelligence from the design file, to the factory, to the returns counter.
- Twin-Grounded VisionIn build
- Traced PerceptionLive on our reference kiosk
- Fused VerdictsMeasured
- Differentiable WearExploring
- Act
Accountable Agents
Agents that do real work inside a business and can always show, check or undo it.
- Source-Bound FiguresIn build
- Version-Bound ApprovalShipped
- Verified EditsShipped
- Curiosity EngineExploring
Underneath everything
- Prove
Measured claims
We set the pass bar before every run, and measure the noise first, so a change smaller than the noise counts as no result.
- Pre-Registered Evaluation
- Egress ProofMeasured
- Own
Owned by construction
We never train on a customer's data for our own purposes, and we never take it. Their open-weight models are trained on hardware in their own building, for them alone. We manage it; it's theirs.
- Self-Checking ReleasesLive pilot on our own box
Proof
If we haven't measured it, we don't say it.
- Resident Inference
A third of the GPU memory
Same open-weight model, Qwen 3.6 35B mixture-of-experts: about 22 GB in our build of the open-source TensorFold engine, against 63 GB reserved by the comparison engine, which is not a measured floor.
- Exact Drafting
Identical output
Faster drafting must reproduce the plain answer token for token. It did in 118 of 118 live replies, and in 160 of 160 across two GPU types.
- Schema-Locked Extraction
94% to 0%
Relationships outside the business's own vocabulary, across 300 real events.
No write-ups yet.
Who we are
Who we are.
Since 2023 we've been building the record of what happens to things, first for physical products and then for whole organisations. Now we build and deploy intelligence compiled from that record.
Where it goes: every product, and every company that builds it, improving from its own record.
- Based in
- London and San Francisco
- Live today
- Seesa, running in customers' buildings, including a FTSE 100 company. Take-back and recycling programmes for global brands.
Contact
Talk to Labs.
Tell us the work you want done. We'll show you how Labs would build and deploy it.