Career Copilot

The live module is private and owner-only. This page publishes only behavior, design decisions, and safety constraints. It includes no candidate snapshots, job listings, employer names, application documents, screening answers, messages, storage paths, access tokens, or operational identifiers. Public assistant sessions remain isolated from private career context.

Each workflow is stored as an agent run with ordered steps, inputs, outputs, statuses, validations, and timestamps. The curator, scout, analyst, matcher, researcher, writer, reviewer, watcher, and coach each have narrow responsibilities. Laravel queues may continue eligible work, but durable state and explicit transitions remain authoritative.

Candidate facts are versioned and approved before supporting a resume or answer. Each fact can carry evidence and an answer policy. Professional wording may be drafted, exact-only facts must be copied exactly, and user-only topics are never answered by AI. Conflicting education, dates, domains, employment, or metrics block snapshot approval.

The source layer accepts manual listings and forwarded alerts, with allowlisted public ATS feeds reserved for a later connector phase. Listing and email bodies are untrusted data, never instructions. Code handles location, work authorization, source rules, exclusions, and minimum requirements before semantic analysis scores the Laravel/Filament and Data/BI tracks. Packages require the configured quality threshold, and close track scores require user selection.

The writer receives an approved snapshot, normalized listing, and allowlisted company research with citations. It drafts a tailored resume, cover letter, screening answers, and resume diff. Every claim resolves to approved facts. The truth reviewer blocks unsupported dates, metrics, credentials, employers, and experience before review.

Filament presents the listing snapshot, fit score, gaps, documents, answers, citations, and diff together. Approval records the exact package version. Any edit changes that version and revokes approval, ensuring submitted content is the content the user reviewed.

The MVP never scrapes, fills, clicks, or submits forms on LinkedIn, Indeed, Glassdoor, or similar services. Approval enables downloads, copy actions, and the official form link. The user completes the third-party form and records submission afterward. Public ATS feeds support discovery, not unrestricted submission.

Private fields, files, session state, and queued career payloads remain in the owner boundary and use encrypted storage or serialization appropriate to their type. Credentials are resolved by trusted services and never enter agent context. Raw forwarded email, unselected listings, and application history have configurable retention periods with export and deletion controls.

Confirmations, interview requests, rejections, and offers become application events. Ambiguous messages require confirmation, while follow-ups stay drafts until approved. The coach builds likely questions, public company research, and STAR stories only from approved evidence.

The semantic request contract is hard-locked to gpt-5.6 with GPT-5.6 Sol, reasoning.effort: max, Pro mode, and store: false. No mini, budget, or lower-quality fallback is allowed. Until external processing is explicitly approved, semantic runs remain paused. If the required model is unavailable, work pauses and retries. A newer flagship is promoted only after passing the full truth, policy, security, and job-quality evaluation suite.

The default cadence is one high-quality package daily, with no forced application below threshold. Analytics focus on source quality, eligibility, score calibration, revisions, interviews, and outcomes. This is governed Laravel product engineering, not a mass-application bot.