Embedded engineering in the AI era: what to ask before coding | Part 1
Begin a connected learning path with a continuing project: define the problem, write testable requirements, review AI suggestions and produce your first engineering brief.
Contents
From beginner to embedded engineer in the AI era · Part 1 of a 200-part learning path
You do not need to know every microcontroller to begin embedded engineering. First, learn to turn a real need into specific, testable behaviour. Then choose circuits, software and tools to support that behaviour. AI can help, but a decision you cannot explain or check is not yet a dependable engineering decision.
An embedded system is a computer inside a device, connected to the physical world for a particular job: it reads inputs, makes decisions and produces outputs. Controlling a light, recording temperature and managing a motor are examples. Software matters, but an incorrect input, unsuitable supply or timing mistake can still defeat the product even when its code compiles.

What should you be able to do by the end?
The goal is not to memorise component names or become dependent on one board. It is to clarify a problem, record assumptions, build a simple solution, investigate faults using evidence and deliver something reproducible. Professional competence also needs practice, feedback and real experience. This series supports that work; reading 200 articles alone is neither a qualification nor a job guarantee.
Our continuing project is a low-voltage room monitor. We begin by defining its behaviour, then progressively add measurement, indication, logging, fault handling and communication. No exact board or sensor has been selected, and this lesson requires no purchase. The initial exercise involves no mains connection, heater control or medical use.
| Parts | Capability we will develop |
|---|---|
| 1–20 | Problem definition, experimental thinking, learning habits, writing and working with AI |
| 21–40 | Voltage, current, power, components and basic measurement |
| 41–60 | C programming, memory, build tools and version control |
| 61–80 | Microcontrollers, I/O, interrupts, timers and timing |
| 81–100 | Sensors, conversion, errors, calibration and interfaces |
| 101–120 | Software architecture, state machines, concurrency and RTOS concepts |
| 121–140 | Debugging, testing, evidence and fault behaviour |
| 141–160 | Power, schematics, PCB design and reviewable prototypes |
| 161–180 | Secure communication, updates, energy use and responsible AI assistance |
| 181–200 | Delivery, manufacturing, documentation, teamwork and a final project |
This is the learning order, not a claim that all parts have already been written or published. Each lesson has a purposeful exercise, and the next begins with a worked answer and common mistakes. Stage-end exercises consolidate the project. If you already know a topic, demonstrate readiness by completing its exercise and explaining your decisions.
Before code: who needs this, and why?
“Build a smart device” is not yet a sufficient problem statement. Imagine someone wants a light to warn that a box lid has been left open. The first question is not which microcontroller to buy. We need to know who sees the light, how long an open lid matters, how closure is detected and what should happen if the sensor disconnects.
For this teaching example, we choose three explicit assumptions: warn after ten continuous seconds open; turn the warning off after detecting closure; and distinguish an invalid input from a closed lid. These are decisions for our hypothetical problem, not measured customer requirements or universal recommendations. A real project must establish them with its user.
| Checkable requirement | Proposed test |
|---|---|
| R1: Do not activate the warning before ten continuous seconds open. | Hold the open input for nine seconds and record the output; examine the ten-second boundary separately. |
| R2: Closing the lid restarts the open-duration count. | Open for six seconds, close, then open for another six; the intervals should not be added without justification. |
| R3: An invalid input has a separate status. | Inject an invalid input; do not silently translate it into “lid closed”. |
No circuit or firmware has been built. The table defines what to build and how to examine it. The NASA systems engineering handbook also emphasises clear, verifiable requirements and their connection to evidence. We apply that principle to a small learning project; our project is not NASA-qualified.
Separate needs, assumptions and solutions
“The user must notice the warning” is a need. “The user always stands in front of the device” is an assumption that may be wrong. “Use a red LED” is a proposed solution. Mixing these categories can produce a well-made circuit for the wrong problem.
For each important decision, record the problem it addresses, its supporting assumption, a simpler alternative and a test that could reveal whether it was suitable. A precise page is more useful than twenty vague pages. When evidence is absent, “not measured yet” is a professional statement, not something to hide.
Why “it works” is insufficient
There are two different questions: does the device meet its written specification, and does that behaviour address the user's need in the intended setting? NASA's discussion of verification and validation distinguishes these perspectives. A warning light can activate exactly as programmed yet be too difficult to notice under room lighting.
Building, simulation, compilation and hardware testing also provide different evidence. A successful compile shows that the build tools accepted the code; it does not establish real response time or correct wiring. This lesson examines examples and requirements documents. It reports no constructed instrument or physical measurement.
Make AI assistance checkable
Instead of asking for an entire finished project, ask AI to find ambiguous requirements, hidden assumptions and failure cases in your draft. Review its suggestions yourself. The NIST generative AI profile discusses confidently incorrect responses, including fabricated reasoning or references. Fluent output therefore does not replace a datasheet, calculation or test.
I have defined a reminder for a box lid left open.
Review the requirements below; do not write firmware yet.
1. Which statements cannot be tested?
2. What assumptions about the input or user are hidden?
3. Suggest normal, boundary and fault cases for each requirement.
Treat unknowns as questions. Do not invent measurements or results.
Before sharing workplace material, check confidentiality rules and permission to use the tool. Do not submit keys, customer information or private schematics merely to obtain a quick answer. Keep AI output labelled as an unverified suggestion until supported by evidence. “Because AI said so” is not a sufficient decision record.
Professional habits start with a small exercise
You do not need to know everything on day one. You do need to ask precise questions, disclose uncertainty and avoid reporting unfinished work as complete. When a disagreement arises, discuss the need and evidence rather than defending a favourite tool. When evidence contradicts a decision, revise the assumption or design; time already spent does not make it correct.
Try allocating a bounded session, perhaps thirty to forty-five minutes, to tonight's exercise. This is a learning suggestion, not a test of intelligence or professional speed. If you get stuck, record the particular term or decision that remains unclear. That note becomes a better question for the next session.
Part 1 exercise: define the room monitor before building it
The raw request is: “I want a device on my room desk that displays temperature status and alerts me when conditions are unsuitable. If measurement fails, it should not display an incorrect number.” The user has not specified thresholds, response time or indication method. Your task is not to select a sensor immediately or generate firmware.
- In no more than five sentences, identify the user, the decision the device supports and what is outside scope.
- Write five questions that genuinely affect the design. Explain what each answer could change.
- Using explicit assumptions of your own, write three numbered, testable requirements. Include units, conditions and fault behaviour where relevant; do not describe assumed numbers as measurements.
- Give each requirement a normal test and at least one boundary or fault test. Record input, expected behaviour and the evidence to collect.
- Ask AI to critique the draft, then record one suggestion you accept and one you reject, with reasons. Without AI access, perform the same review yourself using the checklist above.
Your practical deliverables are a problem brief and a test matrix. Use the English problem-brief template and starter test table. No hardware purchase or assembly is required yet. The lid-reminder examples demonstrate the method; do not copy them into a temperature problem without reasoning.
Self-check: could each requirement produce a pass or fail? Is “invalid measurement” distinct from “acceptable temperature”? Are assumptions and unperformed work visible? Several different answers may be valid; reasoning and checkability are the goal. Part 2 begins with a complete example answer, its decisions and common mistakes, then develops the distinction between a need and a solution. No link is invented for an unpublished lesson.
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.