Wishflip · research & development
The future doesn’t
build itself.
We build the intelligence that builds what’s next — harnesses, orchestration, compute, and the hardware it all runs on.
The future doesn’t build itself.
We build the intelligence that builds what’s next.
Wishflip is a research and development company working on artificial intelligence, compute infrastructure, and the hardware it runs on. We design the systems that make intelligence useful, then we make them behave.
AI is the spark. Engineering is what makes it useful.
A model on its own is a very confident autocomplete. Put it inside a system with tools, memory, boundaries and someone accountable, and it starts doing work you would actually sign off on.
We work across four layers: intelligence, orchestration, infrastructure, and the interface a human touches. Most teams have one of them. The interesting failures happen between them.
A model is powerful. A harness gives it a job.
We build the scaffolding around models — tools, memory, policies, evaluation, and a clear way to stop. The model proposes; the system decides what is allowed to happen.
AI harness engineering
The loop, the tools, the guardrails, the budget.
AvailableAgent design & orchestration
Roles, handoffs, and who is allowed to do what.
AvailableTool & workflow integration
Real systems with real permissions, not demos.
AvailableMemory & context systems
What an agent remembers, and what it is never told twice.
In developmentEvaluation & observability
Scores that mean something on Tuesday, not just launch day.
In developmentHuman approval & escalation
The moments where a person must be in the loop, by design.
AvailableMulti-agent collaboration
Agents that hand off work instead of arguing about it.
In development
We don’t just automate tasks. We teach the orchestra to play together.
Automation usually means one system shouting instructions at another until something breaks. Orchestration means the whole set of systems knows the piece, the tempo, and what to do when a cello walks out.
Business process automation
The repetitive work, removed without removing the judgement.
AvailableAI-assisted operations
Copilots for the people who actually run the thing.
AvailableSystem & API integration
The unglamorous glue that makes everything else possible.
AvailableDocument & knowledge workflows
From a filing cabinet to an answer, with a citation.
In developmentDecision support
Options, trade-offs, and the evidence behind each.
In developmentHuman-in-the-loop processes
Approvals that are real, not theatre.
AvailableCustom operational copilots
Built for your work, not a generic demo.
In development
Intelligence needs somewhere to live. We build it a very nice home.
Architecture, compute, and the operational discipline that keeps it running at 3am. We are as interested in the boring parts — capacity, failure, cost, recovery — as in the benchmark.
AI infrastructure architecture
What to run, where, and what it will cost at scale.
AvailableServer & workstation design
Fits the work, not the datasheet.
AvailableGPU compute environments
Scheduling, isolation, and utilisation you can measure.
In developmentHigh-performance computing
The unglamorous engineering behind fast results.
ResearchDeployment & integration
Into the systems you already have, without a rewrite.
AvailablePrivate, hybrid, and edge
Where the data has to stay, the data stays.
AvailableSupercomputer & cluster concepts
Design studies, not shipped machines.
Concept
From the cloud to the board. From the board to the bold idea.
Software is only half the story. Sensors, microcontrollers, and the strange joy of getting a 300 kilobyte binary to do something useful in the physical world.
Microcontroller & embedded systems
ESP32-class hardware, firmware, and the toolchain around it.
In developmentPC board & electronics development
Schematics, layout, and the first prototype that works.
ResearchHardware–software integration
Where the firmware meets the dashboard.
In developmentEdge AI concepts
Small models, small budgets, real constraints.
ResearchSensor & device integration
Getting data out of the physical world reliably.
In developmentComputer architecture research
How the next layer down might be organised.
ResearchChip technology exploration
Reading, prototyping, and being honest about the gap.
Concept
We distinguish these on purpose. Nothing here claims a fabricated chip, a manufactured device, or a benchmark we have not run.
The best intelligence is the kind that makes people more capable.
Interfaces that behave like instruments rather than oracles. Voice, touch, context, and a clear answer to the question every person asks a new machine: what are you doing, and can I stop you?
Human–computer interaction
Interaction designed for trust, not novelty.
In developmentWearables & companion interfaces
A pendant that belongs to you and answers to you.
In developmentNatural interaction
Voice, gesture, and context that do not need a manual.
In developmentAccessible, understandable AI
Explanations a person can act on.
AvailableHuman agency & privacy
Consent, control, and the right to be left alone.
Available
Have an idea that sounds slightly impossible? Excellent. Those are our favourite kind.
Some of what we build begins as a sketch on a whiteboard and a sentence that starts with “wouldn’t it be funny if”. We keep those sketches, and we tell you which ones are still sketches.
Meet the agents we actually run.
Not a marketplace of imaginary assistants. The systems below are the ones in use inside Wishflip, with their boundaries stated as plainly as their abilities.
The future is a work in progress. Want to help us make it interesting?
Tell us what you are imagining. We will tell you honestly whether it is a weekend, a quarter, or a genuinely hard research problem.
Measured, not claimed
The numbers we can actually show you.
Every figure below was read off our own hardware this week — the same two-core host that serves this page, and the same station toolchain that signs the firmware we ship. If a number is missing, it is because we have not measured it yet.
- 5.5ms
Warm render, heaviest console pagewas 1,217 ms before we removed a subprocess from the request path
- 68ms
Fastest edge response, p50 across routesworst route measured 117 ms, through Cloudflare
- 4.19MB
Signed factory image, byte-verifiedbootloader, partition table and application in one burn
- 76tests
Passing in the device platformincluding a tampered firmware image being refused at flash time
Timed in-process after warm-up, no cache, no CDN. The outlier is a page that compiles nothing and only lists a queue — it is slow because of a query we have not optimised.
- Bootloader + partition table 64 KB
- Application 314 KB
- Free for the next feature 3,718 KB
Verified by offset, not by filename: bootloader at 0x0, partition table at 0x8000, application at 0x10000. Writing the app at 0x0 bricks the board, which is why the signer only ever releases the merged image.
Time to first byte from a real browser through the edge, measured per route. The spread is TLS and geographic distance, not rendering.
The next move
Tell us what you’re
imagining.
We’ll tell you honestly whether it’s a weekend, a quarter, or a genuinely hard research problem. All three are fine.
- AURONIn developmentA wearable pendant that authenticates its owner
- MakerIn developmentSentences into working firmware
- Compute estateAvailableThe infrastructure behind the work
End of the line
You just travelled
the whole internet.
Orbit, atmosphere, exchange, fibre, city, home. Every layer of it was built by somebody, and most of it is still maintained by hand.
>_ traceroute orb-14 → you
satellite 62ms · exchange 3 hops · fibre 1,412km · city 9ms
>_ hops remaining
0
>_ echo $PURPOSE
"build the parts that make people more capable"