Evidence Engineering: Why AI CV Optimization Needs More Than Your CV

When I started building an AI-assisted CV optimization solution, I wasn’t interested in creating another “CV facelift” tool. There are already plenty of those, and all promise nothing short of magic.

Most produce well-polished documents with rich wording. Whether those documents still describe the candidate accurately is a different question.

The objective of my solution evolved during the implementation - from assembling a job-specific CV to increasing the relevance of an existing CV without introducing unrealistic information.

That design objective led to a very different architecture.

Most AI CV tools solve the wrong problem.

A typical AI CV optimizer receives two inputs:

• the current CV;

• the job description.

The model is then expected to:

1. analyse the vacancy;

2. identify missing keywords;

3. assess the candidate;

4. rewrite the CV.

From an AI perspective, this is a creative writing task. The model is expected to bridge the gap between the CV and the vacancy. Unfortunately, that is exactly where hallucinations occur.

A language model doesn’t distinguish between “making the CV more relevant” and “making the candidate more attractive”.

If the job description emphasizes strategic sourcing and the CV mentions procurement activities, many models will happily transform one into the other. The result may read well, but it is no longer a faithful representation of the candidate’s experience.

Separate assessment from editing

My friend, the Business Automation Assistant, was assembled specifically for workflow creation guidance. Together, we decided to split the solution into two completely independent stages:

1. The first stage performs the assessment. It analyses the CV against the vacancy and produces a structured assessment that includes optimization priorities, keyword gaps, evidence coverage, employer priorities and recommendations.

2. The second stage never performs another assessment. Its only inputs are:

   
   

   the original CV;

   the assessment output.

   
   

   The assessment tells the engine what should change. The CV tells the engine what evidence exists. These two responsibilities are never mixed.

   
   

Evidence Engineering: Building the Foundation Before You Apply

Most discussions about AI-assisted job hunting start with prompts. That’s the wrong place to start.

The first step is defining not your job title, but your professional persona. What problems do you solve? What distinguishes you from other procurement professionals? Are you a transformation leader, a category specialist, a commercial strategist, an operating model designer, or a procurement executive? Without that context, AI has no way of deciding which parts of your career deserve prominence.

The second step is defining your career objectives. Every optimization decision depends on where you want to go, not just where you have been. The same experience should be presented differently depending on whether you are pursuing executive leadership, consulting assignments, interim management, academia, or AI-related roles. A technically accurate CV can still be strategically wrong if it points your career in the wrong direction.

The third step is engineering your evidence. A CV contains only a fraction of your professional experience. The real work is extracting and structuring everything that may become relevant in future applications: achievements, procurement categories, transformation programs, negotiations, supplier relationships, technologies, publications, certifications, training, presentations, governance activities, methodologies, and measurable business outcomes. Every piece of evidence must be attributable, reusable, and linked to the roles where it was acquired.

Only after this foundation exists does prompt engineering become useful.

The prompts themselves are only one component of a much larger workflow. They define how employer requirements are analyzed, how evidence is evaluated, how optimization priorities are determined, how unsupported claims are rejected, and how approved changes are transformed into CV edits. The quality of the final application depends far more on this decision logic than on the language model itself.

Evidence Engineering refers to methods to produce accurate and actionable evidence from operational data.

This is why I refer to the process as Evidence Engineering (rather than CV writing. The objective is not to produce a better document. The objective is to build a governed, evidence-based representation of your professional career from which accurate, vacancy-specific applications can be generated consistently and at scale.

AI CV Optimization Workflow: This infographic details a systematic, two-stage approach to refining résumés using AI, emphasizing evidence-based inputs to align users' career goals and experiences with job requirements. It highlights the preparation, assessment, and transformation processes, ensuring that CVs are optimized without inventing or misrepresenting information, maintaining relevance and accuracy.

The AI CV Optimization engine is not allowed to think creatively.

Most prompts tell an AI to improve, strengthen, optimize, or rewrite text.

Those instructions encourage creativity.

My transformation engine is deliberately restrictive.

It is not allowed to:

• invent achievements;

• infer responsibilities;

• strengthen claims;

• introduce technologies;

• create leadership experience;

• complete missing information.

If a proposed change cannot be supported by both the assessment and the existing CV, the change is discarded. No exceptions.

In practice, the engine often produces fewer edits than a conventional AI tool. That is intentional. Omitting a weak edit is preferable to introducing an unsupported claim.

The transformation process is deterministic.

The engine follows a fixed sequence.

1. Read the assessment.

2. Read the current CV.

3. Locate the target section.

4. Locate the existing wording.

5. Verify that the assessment requests the modification.

6. Verify that the replacement is supported by the existing CV.

7. Produce the edit.

8. Otherwise, omit it.

The engine is not rewarded for producing more edits. It is rewarded for producing only defensible edits.

Why this approach matters

Recruiters increasingly use AI to screen applications. Candidates therefore need CVs that align with the vacancy. However, alignment should not come at the expense of accuracy.

Many AI tools optimize for persuasion. My solution optimizes for evidence.

That makes it more conservative than most CV editors. It also makes the output considerably easier to defend during an interview.

Ultimately, the goal is not to produce the most impressive CV. The goal is to produce the strongest CV that remains completely supported by the candidate’s own experience.

Another point is that we have to stop believing in magical tools that work upon a minimum readily available input, i.e., a CV. Any quality process requires quality data to intake. You will have to spend hours, if not days, explaining to your LLM companion what you are capable of, what you know, and what you want to achieve.

Lack of preparation or simply the "magic bullet" approach will get you nothing but a nicely hallucinated custom CV.