Needs versus solutions | Embedded Engineering in the AI Era, Part 2
A worked room-monitor brief, three requirements and six test cases. Distinguish needs, constraints and solutions, then build a reasoned decision log with AI critique.
Contents
The answer to Part 1 starts with defining the room monitor, rather than naming a sensor or board. First identify the user’s decision, the information still missing and the behaviour that could be checked. The worked answer below is one reasoned example. Yours can differ if its assumptions and evaluation criteria are explicit.
This is Part 2 of “Embedded Engineering in the AI Era.” If you missed the exercise, read Part 1 and the original room-monitor request first. We are not selecting components or assembling a circuit yet.
Part 1 answer: a five-sentence problem brief
Our example user works at a room desk. They want to see temperature status near the desk to decide whether to change the room’s conditions. The device must distinguish an invalid measurement from an unsuitable temperature. Heater control, medical use and remote viewing are currently outside scope. The threshold, alert method and measurement quality still need clarification with the user.
Those five sentences establish a boundary. Someone at a desk and someone in another building may need different interfaces. Excluding heater control does not mean it is impossible; it prevents adding another responsibility to this particular prototype without a reason.
Five questions that change design decisions
- What condition is unsuitable, and what action will the user take? The answer determines the threshold and meaning of an alert.
- What measurement error is acceptable, and what reference will we compare against? This affects sensor selection, mounting and evaluation; display digits do not answer it.
- How quickly must a change become visible, and how old may data be? This changes sample intervals and stale-data behaviour.
- Will the user always see the display, or is an audible alert needed? This identifies the output interface and usage conditions.
- What power and communication arrangements are permitted? A fixed-power local display and a battery-powered remotely viewed device have different needs.
A useful question has a consequence: you can explain which decision its answer changes. You do not need every answer immediately. Record an unknown so it does not silently become a fact in the design.
Three example requirements with explicit assumptions
For the exercise, assume the user is nearby, accepts a fictional threshold of 26.0 °C, expects the relevant output to change within one second of an event, and considers data stale after five seconds. These numbers are neither comfort recommendations nor experimental results. The exercise input is a temperature sample accompanied by a valid/invalid flag. Establishing actual sensor validity belongs to later project stages.
- R1 — Display: After receiving a valid, fresh sample, the device shall display its value with one decimal place and the °C unit within one second.
- R2 — Status: For that valid, fresh sample, a value below 26.0 shall produce “normal,” while a value equal to or above 26.0 shall produce “high.” The status shall change within one second of sample receipt.
- R3 — Unreliable data: After an invalid input, or when five seconds have elapsed since receipt of the latest valid sample, the device shall show “invalid data” and hide the numeric temperature within one second. The same state shall apply from startup until the first valid sample.
R3 takes priority over R1 and R2: an old number must not continue to appear “normal.” A fresh, valid sample restores R1 and R2. Here, data age is measured since receipt of the latest valid sample, not since the program loop last ran. Sensor quality, permitted input range and final measurement accuracy are not yet approved specifications. These three requirements are not the complete product specification.
Six test cases: an expectation is not a result
The table is a test plan; actual result fields in the downloadable file are blank. For timing, capture both the input event and the output change. Seeing the right number without knowing when it appeared does not fully verify R1.
| Case | Input or boundary | Expectation and evidence |
|---|---|---|
| T1 / R1 and R3 normal | Valid sample 21.4 | Display 21.4 °C without an invalid label within one second; record value, status and both timestamps. |
| T2 / R1 boundary | Response close to one second | 1.00 s is within the limit; 1.01 s is not. Record actual timing and measurement method. |
| T3 / R2 normal | Valid sample 25.9 | Normal within one second; record input value and status text. |
| T4 / R2 boundary | Valid sample 26.0 | High within one second; check the “equal to or above” boundary. |
| T5 / R3 fault | Invalid input, even with value 25.0 | Invalid data and hidden number within one second; a plausible value must not conceal the fault. |
| T6 / R3 boundary | Latest valid sample at t=0; age 4.9 then 5.0 s | Fresh at 4.9; stale at 5.0. Invalid status must be visible by t=6.0. |
The six-case worked test table contains inputs, expectations and blank observation/result fields. These boundaries belong to the defined model. In a real test, account for the timing instrument’s resolution and uncertainty when judging the result. Startup and fault-recovery cases will extend this initial set.
Accepting and rejecting a critique, with reasons
Consider two hypothetical critique suggestions. The first says, “Separate invalid data from high temperature.” We accept it because a lost measurement and a high reading are different conditions; R3 expresses that distinction. The second says, “Require an ESP32 and cloud uploads immediately.” We reject it for now because remote viewing is absent from this example brief. These are teaching examples of a decision record, not feedback received from you or reported experimental outcomes.
Common Part 1 mistakes include choosing a sensor instead of describing the problem, saying “fast and accurate” without criteria, omitting the equality boundary and filling the result column before testing. Another is promoting an exercise assumption to a confirmed user request. Correct this by recording each number’s or decision’s origin: user statement, existing constraint, provisional assumption or observed result.

Needs, requirements and solutions
A need is the user’s problem or desired outcome: “Understand a change in temperature conditions while at my desk.” A checkable requirement states the system’s behaviour and conditions, such as R2’s threshold and deadline. A solution is a way to deliver that behaviour: a display, light, buzzer or phone message. One need can admit several solutions, each with costs and limitations.
This distinction aligns with the progression from expectations through technical requirements to solutions in NASA’s systems engineering guidance. A small project does not require the administration of a space programme. A clear note explaining decisions is sufficient here. The useful principle is retaining the connection between a problem and the decision made to solve it.
A component name is not automatically a mistake. If a client actually requires using a board already in inventory, that board is a constraint; record its source and rationale. If it is merely our first idea, do not present it as the user’s requirement. The stakeholder expectations guide likewise connects understanding needs, constraints and assumptions with discussion and documentation.
Compare solutions against a reason
For this example, a local display can be read without a network dependency, but the user must be nearby. A phone message can support remote viewing, but its communication path and outage behaviour need definition. The brief currently assumes a user at the desk, so mandating networking lacks sufficient justification. If remote viewing later becomes a user need, update the problem and tests along with that change.
Keep a simple chain: understanding temperature status leads to R2; R2 needs a distinguishable output; candidate outputs are compared; T3 and T4 check the chosen candidate’s behaviour. That connection helps identify the need behind a feature. A feature without a need can simply add time, consumption and debugging work.
Use AI to critique reasoning, not invent user wishes
Ask the tool to expose ambiguity and hidden assumptions, and to identify the brief statement supporting each suggestion. For example: “Which need makes networking necessary? Without evidence, express it as a question or option.” Fluent output does not replace user input or technical evidence. Section 2.2 of NIST’s Generative AI Profile describes confidently presented incorrect content.
In a team, connect disagreement to a decision and reason: “Keep this option open until remote viewing is clarified,” rather than labelling a person or tool. Record the open question, who will follow it up and the evidence needed. Critiquing this public example requires no private workplace project information.
Part 2 exercise: create a decision log
Continue the room monitor. Classify these six statements using their context, recording a reason and the remaining question for each:
- “I want to notice a change in temperature status while at my desk.”
- “The display must update quickly.”
- “I suggest using an ESP32.” The speaker is the designer and has stated no contractual constraint.
- “If no valid sample arrives for five seconds, stop showing the previous number.” The number remains a provisional team assumption.
- “Everything must go on the internet.” No remote viewer has yet been identified.
- “The installation location only has the existing five-volt USB output.” This time, the client has confirmed the site constraint.
Then turn the ambiguous second statement into a checkable requirement. Connect one real need to two candidate solutions and propose two comparison criteria with reasons. Give the new requirement one normal test and one boundary or fault test. If you obtain AI critique, document one acceptance and one rejection; without a tool, perform the same review using this lesson’s questions.
The deliverable is an English decision log. Self-check: are the sources of constraints and numbers visible? Does “quickly” still lack units? Is each solution connected to a need? Does each test specify an input and expectation? Part 3 will open with a reasoned answer to these exact six statements and the candidate comparison, before examining users and operating context more closely.
Question and answer article
Ask your technical question
Questions and answers are published after moderation. Authors can see their own pending drafts.
No confirmed questions have been registered for this article yet.