Two to six weeks, real parts under real conditions, and a written answer — which is sometimes that the problem cannot be solved as posed. The purpose is to move the expensive discovery from month seven to week two, and a study that only ever returns encouraging answers is not doing that job.
The business complaint is turned into a measurable task with a defined cost of error. This alone sometimes changes the project, because the stated problem and the expensive problem are often not the same.
Real parts or real signals, under real conditions, including the rare cases the system exists to catch. Where those do not exist in sufficient number, that constraint is established now rather than discovered during training.
Several configurations compared empirically. For vision this means several illumination geometries on the same defective parts; for condition monitoring, several sensor placements and sampling rates.
Classical methods first, because if a deterministic approach suffices it is the better answer. Learned models where appearance or behaviour varies beyond what a rule can express.
Latency and power measured on representative target hardware, not estimated from operation counts, which correlate poorly with wall-clock time on embedded targets.
Findings, recommended sensing configuration, achievable accuracy with its confidence, deployment outline, cost and risk. Including, where it applies, a recommendation not to proceed.
A study that concludes the problem cannot be solved as posed has usually saved a multiple of its own cost, and it almost always comes with an alternative: change the lighting, add a different sensor, move the inspection point, instrument the process differently, or accept that this particular defect is a process problem rather than a detection problem.
The alternative is a programme that runs for two quarters before the same conclusion arrives, by which point hardware has been bought and expectations have been set inside the customer's organisation.
This is also why a supplier who has never returned a negative feasibility result is worth asking about. Either the work is unusually easy, or the study is a formality.
Typically two to six weeks, depending on how much real data already exists and how difficult the capture conditions are. Where the target event is rare, the limiting factor is usually how long it takes to observe enough of them.
That is a successful outcome and generally saves a multiple of the study's cost. It normally comes with an alternative — a sensing change, a different measurement point, or a process change that removes the need for detection.
A written document: findings with supporting evidence, the recommended sensing configuration, achievable accuracy with its confidence, a deployment outline, cost and remaining risks. Not a slide review.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.