Anton.FindJobHelper.Generation 1.0.0

FindJobHelper

This is a tool used to generate a CV by matching the given tags to your experience database.

Development tests

The LaTeX-dependent tests use a minimal, repository-local TeX Live installation. Set it up once before running the complete test suite:

./scripts/setup-tests.sh
dotnet test FindJobHelper.slnx

The toolchain and Liberation fonts are installed beneath the ignored .tools/ directory. Test assemblies discover that installation automatically. Without it, tests that render or measure LaTeX fail; tests that do not use LaTeX remain available.

Usage

Basic idea

// config.json
{
    "requiredTags": [
        { "name": ".NET", "weight": 1.0 },
        { "name": "Microservices", "weight": 1.0 },
    ],
    "sections": {
        "workExperience": {
            // Only include a single bullet into the CV
            "itemBudget": 1
        }
    }
    // Includes more configuration here, like the
    // listed skills & technologies, page layout, matching parameters.
}

// Experience Database (some code omitted for brevity)
// Must be compiled into a dll to be used by the generator.
var builder = new ExperienceDatabaseBuilder();
var company = builder.Place("Example Company");
builder.Job(job =>
{
    job.Title("Example Software Engineer");
    job.Place(company);
    job.DateRange(DateRange.Completed(new(2023, 1), new(2024, 12)));
    
    // Good match, because of the tag match
    job.Item(item =>
    {
        item.Text($"Built a fictional .NET service for example users.");
        item.Tag(Tags.DotNet, 10);
    });
    
    // Bad match, no matching tags
    job.Item(item =>
    {
        item.Text($"Wrote app in C++");
        item.Tag(Tags.Cpp, 10);
    });
});

Then you run the generator:

dotnet find-job-helper --config config.json --experience-database compiled_database.dll

And get a FirstLast.pdf rendered CV, containing the best-matching experiences according to your configuration.

In this case, the .NET bullet is going to appear in the output, while the C++ one is going to be skipped. This is because we set itemBudget to 1. In principle, without that setting, it takes as many items as fit the page, ensuring the best matches are selected.

Setting up the workspace

1. Install and instantiate the templates package:

   dotnet new install Anton.FindJobHelper.Templates
   dotnet new findjob-workspace -n ExampleWorkspace
   cd ExampleWorkspace
   dotnet tool restore
   

2. Install Latex and the default fonts using either the installation script provided with the template, or manually. Default fonts can be overwritten, you don't necessarily have to install them.

   ./scripts/setup-latex.sh
   source "$HOME/.local/share/findjobhelper/texlive/2026/findjobhelper-env.sh"
   ./scripts/setup-latex.sh --check
   

   The installation script has only been tested on Ubuntu.

3. Inspect and edit the experience database at src/FindJobWorkspace.Provider/ExperienceDatabaseProvider.cs. Inspect and edit the run script (run.sh or run.ps1) to specify your personal data and overwrite default fonts.

Font configuration

The generator accepts separate family and scale settings for its main, sans-serif, and monospaced fonts. A command-line option takes precedence over its environment variable.

The font-size options are dimensionless positive, finite scale factors. The generator passes them to fontspec as Scale=<value>; they are not point sizes.

To generate a CV, do the following:

1. Make a config by using dotnet find-job-helper new-config in a new directory for the current run. Edit the config.json to fit the tags of the target job position.

2. Run run.sh or run.ps1 from the directory with the config.

Usage with agents

I personally mostly use this with an agent. Just tell it to generate a CV from the workspace folder and give it a link to the job posting.

If your experience database is not initialized yet, and you want a starting point, tell the agent to scan your GitHub repositories or your contributions in repositories at work, and record the experiences and all new relevant tags in the database. I recommend reviewing each experience the agent adds, carefully evaluating the right values for the tag weights, and rephrasing the sentences until they are sound. Also, the agent may decide to include statistics that it found important, but which are actually irrelevant, so be careful of those.

Hardcoded defaults

• User parameters are configured per invocation via environment variables or user-secrets. A config can override the profession and the displayed link order for one CV.
• Header links use PersonalInfo__GitHub, PersonalInfo__LinkedIn, PersonalInfo__YouTube, and PersonalInfo__Portfolio. Missing links are omitted in the default GitHub, LinkedIn, YouTube, Portfolio order.
• If education is included, each education experience is forced to appear, unless no bullets pass an earlier filter.
• All job experiences are included bare even if they have no matching bullets.
• Personal projects are not included unless they have at least one selected bullet.

Configuration parameters influencing the matching algorithm

Below is an AI-generated summary of the parameters specified in config.json.

Search parameters

selection.default.directMatchBoost is inherited when a section omits the property. A section value overrides it, and an explicit 0 disables the inherited boost. This nullable inheritance rule applies only to directMatchBoost; all existing selection properties keep their previous overlay behavior. Omitting the property everywhere resolves it to 0 and preserves existing clean-CV selection.

MMR parameters

MMR (Maximal Marginal Relevance) re-ranks the eligible bullets after tag matching. It rewards relevance while discouraging bullets that repeat the same tag profile. This helps a CV cover several requirements instead of spending most of its space on several nearly identical bullets.

For a candidate, the implementation calculates:

MMR score =
    relevanceWeight × normalizedRelevance
  - (1 - relevanceWeight) × maxSimilarity
  - saturationPenalty × saturation

Where:

• normalizedRelevance is the candidate's fully adjusted, pre-MMR relevance divided by the highest such relevance among all candidates. When several requirements reach the same experience tag, base relevance uses only the largest effective coefficient, so aliases and overlapping relation paths do not inflate it.
• maxSimilarity is the highest cosine similarity between the candidate's explicit requirement-coverage vector and any already selected bullet's vector.
• saturation is a weighted penalty for explicit requirements that have already appeared in at least saturationQuota selected bullets.

Configured tags that are bidirectional full-overlap aliases form one requirement group. The first-declared database tag is its canonical name, and the group's configured weight is the maximum alias weight. For example, configuring both C# and .NET at 1.5 creates one canonical .NET requirement with weight 1.5, not 3.

Every matched experience tag retains all explicit requirements that reached it. Its largest effective coefficient contributes to base relevance. A separate direct coefficient exists only for an exact configured tag or a bidirectional full-overlap alias, and uses the largest configured weight in that alias group:

baseTagContribution =
    evidence score × maximum effective requirement coefficient

directContribution =
    evidence score × maximum direct-or-alias coefficient

directBonus =
    max(0, directContribution) × directMatchBoost

tagRelevance = baseTagContribution + directBonus

coverage for an explicit requirement =
    sum of its effective contributions across matched experience tags

Partial, one-way, and transitive relations are indirect and receive no direct bonus. Each bullet tag earns its own bonus; the aggregate bullet score is never multiplied once per matching tag. If a bullet tag is a stronger indirect match for one requirement and a weaker direct match for another, only the weaker direct contribution is boosted. For example, base relevance 10, direct contribution 4, and directMatchBoost: 0.5 produce 10 + 4 × 0.5 = 12, not 15.

After tag relevance is complete, recency is added from unboosted base relevance:

appliedRecencyBoost = configuredRecencyBoost × normalizedRecency
recencyBonus = max(0, baseRelevance) × appliedRecencyBoost

adjustedPreMmrRelevance =
    baseRelevance + directBonus + recencyBonus

The bonuses are additive rather than compounded. With base relevance 10, direct contribution 4, direct boost 0.5, and applied recency boost 0.25, the result entering MMR is 10 + 2 + 2.5 = 14.5, not (10 + 2) × 1.25 = 15.

This means an indirect Unity or Game Programming match reached from .NET still occupies the .NET MMR dimension. Transitive intermediate tags do not become MMR dimensions unless they were explicitly configured as requirements. Direct and recency bonuses do not change requirement coverage, cosine-similarity vectors, saturation proportions, or selected-requirement counts. MMR subtracts its similarity and saturation terms only after both positive relevance bonuses have been applied; no later conversion rescales the signed result.

The configuration requires all three MMR parameters:

The standard starting values are:

{
  "relevanceWeight": 0.72,
  "saturationQuota": 2,
  "saturationPenalty": 0.18
}

With saturationQuota: 2, the first two selected bullets covering a requirement incur no saturation penalty from that requirement. A third candidate covering it is penalized once, a fourth twice, and so on. Each selected bullet increments a canonical requirement at most once. A candidate's contribution is proportional to how much of its total requirement coverage comes from that repeated requirement.

The similarity and saturation penalties are global across the selected experience bullets, not reset for each section. Selection stops normally when no remaining candidate has a positive MMR score, unless a section minimum still needs to be filled.

Tuning guide

• Too many similar bullets: lower relevanceWeight, lower saturationQuota, or raise saturationPenalty.
• Important requirements disappear: raise their requiredTags weights or raise relevanceWeight.
• One section dominates: lower that section's itemBudget.
• Older but more relevant work is displaced: reduce recencyBoost.
• Weakly related bullets appear: raise scoreLowerBound. Tune this carefully because it is a hard pre-MMR filter.
• Use --debug after each change to inspect aggregate and per-bullet scores directly in the generated CV.

Contributing

Note that more than 50% of this repo is AI-generated slop, including 90% of the tests. I did the foundation, including the initial experience database, the basic tag matching algorithm, the tags database, the rich text module, and the initial latex rendering module, while the rest is slop which I did not verify very carefully.