Algorithm engineers turn model failures into experiments someone can rerun. AI product managers turn a user task into a workflow with acceptance criteria. If an assistant applies an outdated rule, the engineer needs to show where retrieval or generation went wrong. The product manager needs to decide what users should see, whom to contact, and when the system should hold back an answer.
Compare the work before choosing
Find two graduate job postings from companies you're considering, then try the two deliverables below. Use the FermatMind 94-question career assessment → if you'd like to clarify which work activities interest you.
## Key takeaways
- Algorithm engineering. Establish a baseline, control variables, and investigate failures.
- AI product management. Define the task, acceptance criteria, and handoff to a person.
- Graduate recruiting. Check the role's specialty, degree requirements, and eligible graduating class. Use a portfolio to address gaps.
One problem, two deliverables
| Compare | Algorithm engineer | AI product manager |
|---|---|---|
| Focus | Find where a wrong answer originated | Give users a path to complete the task |
| Skills and output | Control variables and document reproducible runs | Map the workflow and write acceptance criteria |
| First exercise | Rerun a case that confuses old and current rules | Write a handoff test for missing evidence |
How do the requirements differ in two Beijing graduate postings from one company?
Check the specialty as well as the title. In these Beijing postings, an embodied AI role asks for a graduate degree and simulation skills. An AI product role looks at education, workflow design, and evaluation.
| Beijing campus recruiting example | Deepexi: large-model algorithm engineer (VLA) | Deepexi: AI product manager |
|---|---|---|
| Eligible class and education | Master's or PhD graduates from the 2026 or 2027 class | 2027 graduates with a bachelor's degree or higher |
| Subject and technical background | Computer science, AI, automation, or similar. Python, PyTorch, Linux, and at least one simulation environment | Computer science, AI, software engineering, mathematics, or related. Understanding of large-model limitations, vector databases, and knowledge retrieval |
| Deliverables | Model baselines, evaluations, and replication in simulation or on real hardware | Requirements, prototypes, acceptance criteria, and reviews |
| Pay in these postings | No comparable figure disclosed | No comparable figure disclosed |
See the company's public algorithm position and product position. Embodied models and enterprise knowledge assistants solve very different tasks. The embodied AI requirements apply to this example. Product roles can focus on applications, models, or platforms. When comparing experienced-hire jobs, record city, experience, level, pay period, and the number of pay months separately. If one side lacks a comparable vacancy, leave it blank. One salary for an experienced hire doesn't establish a pay gap for graduates.
A knowledge assistant answers incorrectly. What does each role investigate?
For the same incorrect answer, the algorithm engineer reproduces how the documents or model led to the error. The AI product manager assesses the user impact and defines how the system should clarify the request and recover.
In this editorial exercise, a user asks for the booking deadline at South Campus. The assistant retrieves a North Campus rule and an expired notice for South Campus, then promises a Friday 17:00 deadline based on the expired notice. Give both roles the question, the retrieved document versions, the answer, and the policy owner's explanation. Engineering determines whether the error came from retrieval, version selection, or generation. Product specifies what happens if the valid rule is still missing: withhold the deadline, clarify which campus the user means, or show an authorized contact.
Engineering delivers an experiment with interpretable results
Freeze the generator, prompts, files, and scoring rules. Compare keyword search with keyword plus semantic search. The synthetic sample has 24 questions: 16 with supporting rules and 8 with insufficient or conflicting evidence. Keep tuning questions separate from acceptance questions.
| Observation | A: Keywords | B: Keywords and semantic search | What both roles can infer |
|---|---|---|---|
| Correct version and clause retrieved for answerable questions | 13/16 | 15/16 | Evidence improved on two questions |
| Appropriate clarification or human handoff when evidence was insufficient | 6/8 | 5/8 | One more response failed the requirement |
The two rows use different denominators and measure different outcomes. Treat 15/16 as end-to-end accuracy and you can miss answers that have the right document but no supporting citation. Microsoft's RAG evaluation documentation separates retrieval and aspects of the generated result. Give a colleague one failure and the intermediate evidence for both configurations, and see if they can locate the difference.
Product writes acceptance criteria for each evidence state
| What the assistant has | What users see | Who takes over |
|---|---|---|
| Current rule and known campus | Deadline and current clause | QA checks the answer against the rule |
| Current rules but campus unknown | Ask the user to choose a campus | User clarifies and the assistant checks again |
| Only an expired notice, or active rules conflict | Withhold the deadline and offer the duty contact | Policy owner confirms which version applies, engineering reruns it, and product checks the next step |
The exercise requires all 8 insufficient-evidence questions to follow the correct branch. A misses two and B misses three, so neither is approved for automatic answers yet. Before retesting South Campus, product asks the policy owner to confirm the current rule. While confirmation is pending, users get the duty contact. Engineering and QA can still reproduce the error using the exercise data. Once the owner confirms the rule, engineering reruns the case, QA checks the deadline, source, and contact option, and product decides whether to enable automatic answers. Keep any tuning examples for regression testing and add fresh questions. Passing all 8 checks meets this exercise's acceptance rule. The project assesses cost, latency, and release authority separately.
Related reading
If you're planning experiments and portfolio work, compare AI and computing courses against real tasks →.
What should a beginner or career changer build for a first portfolio project?
For an algorithm engineering application, build a single-variable experiment someone can rerun. For AI product management, write criteria business and QA can use. Add specialty requirements from the target role.
- Algorithm engineering. Design 10 synthetic policy questions. Tune with 6 and run one blind check with the other 4. Deliver baseline results, retrieved material, failure analysis, and rerun instructions. For embodied AI roles, separately check simulation, controls, and hardware requirements. This exercise illustrates one evaluation method.
- Product management. With their consent, watch two colleagues use the old process to find a rule and note where they get stuck. Write a one-page workflow with examples for answering from evidence, clarifying a question, and handing off to a person. Have someone rehearse it using only the criteria. If they don't know what the user should see, clarify the conditions.
- Choosing a role. Both jobs involve metrics and working with others. Note which problem you'd willingly tackle three times and which concrete feedback helps you make the next revision. Use real vacancies to check requirements.
Frequently asked questions
Do AI product managers have to train large models?
It depends on the role. Application, evaluation, platform, and model teams have different responsibilities. See whether the vacancy calls for experiments, code, model training, or workflow acceptance.
Do algorithm engineers have to have a master's degree?
The embodied AI graduate role here asks for master's or PhD graduates. Its education requirements apply to that job. Check the requirements of each position before applying.
Can I apply for both jobs with the same AI project?
Yes. Prepare a separate experiment report and product acceptance brief. Identify your responsibilities, the samples you used, and what hasn't been verified. Provide a way to reproduce each deliverable.
Explore further
- Clarify your preferences. Compare debugging, problem definition, and cross-team release work. Open FermatMind's 94-question career assessment →
- Read real job postings. Mark duties, evidence, and open questions in two descriptions. See how to compare job postings →
- Check what you'll get from a course. Look for assignments that get specific, usable feedback. See how to evaluate a course for a career change →
Editorial review and methods
FermatMind Editorial used two Deepexi postings for the Beijing graduate job requirements. The knowledge assistant and its results are a synthetic example of retrieval, generation, and user acceptance checks. We did not obtain matched pay samples for the two jobs and do not rank their earnings.


