01
Landscape
Who is actually operating in a space, what they have shipped, what they charge, where the gaps are. The sweep that would take an analyst six weeks and be out of date in eight.
- Mapped landscape
- Sourced claims
- Gap analysis
We do theresearch.
Companies come to us with a question they do not have time to answer properly. We answer it at machine scale, checked by hand & where the answer turns out to be a product, we build that too.
- Landscape
- Technical
- Feasibility
- Build
The problem
Most researchdies as a slide.
The usual outcome of a research engagement is a deck. Sixty pages, delivered on a Thursday, skimmed once and never opened again, because it stopped at exactly the point where somebody would have had to do something.
The gap is rarely knowledge. It is that nobody owns the step between finding out and acting on it, and by the time that is resolved the research has gone stale.
What we take on
Four kinds of question.
02
Technical
Whether a technology does what its vendors say. We read the papers, the changelogs and the issue trackers, then build the smallest thing that would prove or disprove the claim.
- Evidence review
- Working spike
- Risk register
03
Feasibility
Whether a specific idea is buildable at your constraints - cost, latency, accuracy, regulation & what the honest version looks like once those bite.
- Constraint model
- Prototype
- Cost and latency budget
04
Build
The part most research stops short of. Where the answer is worth acting on, the same team that found it engineers the thing & stays to operate it.
- Production system
- Handover
- Ongoing maintenance
How it runs
Machine scale.Human verdict.
AI is why we can read ten thousand sources instead of two hundred. It is not why you should believe the conclusion - that part is still done by a person who can be argued with.
01
Frame
The first days go on turning a vague brief into a question with a checkable answer. Most engagements are won or lost here.
02
Gather
Retrieval across filings, papers, patents, documentation, repositories, forums and archives at a volume no reading team would attempt.
03
Verify
Every claim that matters is traced back to a primary source by a person. Anything that cannot be traced is reported as unverified rather than quietly rounded up.
04
Deliver
A written answer, the evidence behind it, and what we would do next. Not a deck built to survive a meeting.
05
Build
Where there is something to build, we build it. That is the difference between a report and a result.
What we will not do
What we will not do
- Present a model’s summary as a verified finding.
- Report a number we cannot trace to a source you can open.
- Fill a page count. If the honest answer is four pages, it is four pages.
- Tell you an idea is feasible because you are paying us to find out that it is.
Open notes
The research we do for ourselves.
A share of our time goes to work with no client attached. It is where the methods get tested before anyone pays for them, and it is published in full.
- 001Retrieval over versioned corporaWhat retrieval should do when the correct answer changed in March and the old answer is still sitting in the index.6 min
- 002Judging answers without a ground truthBenchmarks work where a correct string exists. Most questions worth asking a model do not have one.6 min
- 003Small models, narrow jobsHow small a model can get before a well-defined production task stops working — and what that saves.5 min
- 004Autonomous survey systemsAerial survey produces enormous quantities of data and very little of it ever becomes a decision.5 min
- 005Latency budgets for agentic systemsA system that thinks for forty seconds is a different product from one that answers in two.6 min