Independent project · Customized private fork
Duplicate postings, unreliable source checks, changed resumes, and uncertain submissions can create incorrect records or repeated work. I customized a Python-based workflow to make the next action depend on the evidence available.
An independent customization and maintenance project based on Mads Lorentzen's AI Job Search, which supplied the original framework and commands.
My contribution
Workflow diagram · Human review at consequential steps
dot is the AI assistant coordinating my requests and local project work. The local workflow handles checks and records; I review decisions and authorize submissions. Hermes reads the existing state for reports and audits.
Set the task, preferences, and constraints.
Clarify missing details, pass constraints and evidence to local tasks, and bring results back for review.
Normalize and deduplicate postings, retain snapshots, and verify employer or ATS availability and hard-gate evidence before ranking.
Evaluate, draft, review, revise, and verify against supported candidate evidence. Prepared materials do not authorize submission.
Review the materials and answers. Approve consequential consent, certification, and the final submission.
Complete authorized actions and any manual or account gates. Pause when a portal requires my action.
Retain success evidence for the same job and account. An uncertain result stays unresolved and requires a status check before retry.
Match the exact job, tracker row, and retained archive. Preview the update, then save only the supported status and artifacts.
An ambiguous result returns to a status check and human review. It is not recorded as applied. Read the uncertain-submission example.
Workflow controls
These checks connect job identity, posting evidence, resume versions, and application status. Each determines what can proceed and what needs review.
Compare job identities and combine matching discovery records, retaining source references and conflicting details.
Next action: Review conflicts before treating the links as one opening; keep ambiguous roles separate.
Keep the posting snapshot, source, and check time together. A blocked or malformed response leaves the check unresolved; it does not establish that the job is closed.
Next action: Inspect an authoritative source before treating the posting as verified.
Check PDF page count and extracted text, then review layout and factual claims. A changed resume or relevant evidence input invalidates the affected prior verification.
Next action: Repeat the affected checks and human review before using that version.
Prepared materials and submission attempts are not confirmed applications. Require success evidence before marking a role applied.
Next action: If the outcome is uncertain, check portal status before deciding whether to submit again. Preview tracker changes before saving.
Synthetic illustration · Control 04
This invented scenario illustrates the application-status control. It is not an execution log or a real application record.
Retain the success evidence, record confirmation, and review the tracker update. Do not submit again.
Keep the submission uncertain. Pause for human review and obtain an authoritative status check before considering another attempt.
Record that evidence before resetting the submission state. Resume preparation and obtain approval for any new submission.
Maintenance and validation
When maintaining resume verification, a key question is whether an earlier result still applies after an edit. The check compares the current inputs with those recorded during verification. Changed inputs reopen the affected checks, so an old result cannot silently cover a new version.
Executed validation: On September 19, 2026, three existing regression tests passed using temporary synthetic records: verification invalidation after input changes, recovery from an uncertain submission, and rejection of confirmation without success evidence.
On September 28, 2026, focused follow-up tests also passed using synthetic records, including planned-versus-sent handling, duplicate message records, and contact-date validation.
These are focused checks of the local implementation, not live portal submissions or a full-workflow test. Local change history and current code support the customizations described above.
Related work
See the Excel and Tableau reporting project and academic ServiceNow portal for different examples of organizing information and checking its quality.