CareerWise

Judgment before output

Most career tools start with the resume. CareerWise enforces a different sequence: before anything is written, you need a clear model of what you're working with. Your skills, your constraints, the role, the market. Understanding precedes optimization. Match analysis sits at the center of this. It scores fit across strengths, gaps, and risks, but it also surfaces the conclusion no career tool volunteers: sometimes the right answer is not to apply. A blocking skill gap, a seniority mismatch, a role that conflicts with your stated constraints. CareerWise will say so. The goal is accuracy under pressure, not application volume. Materials come after the reasoning is done. Resume, cover letter, interview prep, pitch, each grounded in the analysis that preceded it. Preparation artifacts, not persuasion theater. You don't feel like you're improvising. You feel briefed.

A career tool where AI suggests, but only you decide what becomes permanent. Your data stays local, and nothing is saved without approval.

The research workspace

Three panels: sources on the left, a research workspace in the center, an output panel on the right. Drop in your resume, a portfolio summary, job descriptions, company research, anything. Sources are tagged and grouped; job postings land in Job Information, your own documents in Your Profile. The center panel is a persistent research log. Ask the AI anything, attach files mid-conversation, or trigger structured analysis. Nothing leaves your device.

Profile extraction and match analysis

Profile Analysis reads across all your uploaded documents and extracts a structured career record: professional summary, tagged skills, experience bullets. Match Analysis then scores how well your profile aligns with the job by skill overlap, seniority fit, and requirement gaps. Both outputs are read-only in the session layer. Nothing is written to your permanent record without deliberate approval. The AI can suggest edits to your profile; you decide what sticks.

Application package

Once sources are loaded and a profile is active, the output panel generates the full application package in one pass: resume, cover letter, interview prep, skills matrix, and a "Why You" pitch. Each document is tailored to the specific job, not a generic template. The interview prep generates both key talking points and suggested Q&A pairs grounded in your actual experience. The pitch generates multiple variants (phone screen, final round) so you have the right register for each stage.

Atomic skills and the ledger

Your experience gets broken down into individual atomic skills, each with a label, a category (engineering, product, leadership, AI, and others), and the specific sentences from your own documents that prove you have it. The AI extracts these from whatever you upload, but they don't enter your record automatically. They land in a pending queue first. You approve or reject each one. Only approved skills flow into any generated output. Visibility is a separate control from approval. A skill can be approved but hidden, preserved for roles where it's relevant, excluded from the current application. This lets you maintain one honest record and compose different presentations of it without ever inventing or duplicating claims. Every action against the ledger is logged: extractions, approvals, rejections, updates, each with a timestamp and the payload that triggered it. The log is append-only and stored locally as plain JSONL. At any point you can read exactly what the system extracted, what you accepted, what you turned down, and when. Nothing about your career record is opaque.

Privacy pipeline

Every AI request passes through a two-tier anonymization pipeline before leaving the machine. Tier A is always on: emails, URLs, phone numbers, and API tokens are stripped and replaced with placeholders before any text is sent to a model. It cannot be disabled. Tier B is opt-in: company names and people names, detected using a dictionary built from your own workspace documents. The mapping between real values and placeholders is stored locally and never transmitted. The same resume can be analyzed by a cloud model without the model ever seeing your employer's name, your manager's name, or your contact details. For people navigating sensitive transitions (still employed, under NDA, searching quietly) this isn't a checkbox. It's the condition that makes the tool usable.

Pipeline and model routing

Five agents run in sequence: job analysis, profile analysis, match analysis, research, and content generation. Each stage is independently routable to a different model. Routing is not arbitrary. It's benchmarked: the same inputs run through Claude, GPT, Gemini, Mistral, Groq, and Llama side-by-side with structured output schemas and diff-scored results. The routing table that ships is the one that passed the benchmark. Final cost reduction: 69% versus routing everything through the top model.