CareerVector product thesis

CareerVector is a private career-intelligence workspace connected to a collectively maintained, real-time model of the job market. It finds opportunities, explains fit and builds truthful applications using the best intelligence available, without requiring the user to understand the machinery underneath.

CareerVector is not primarily an AI CV generator, a job tracker or an LLM frontend. Its promise is to continuously understand one person's possibilities and the live job market, then help that person choose and pursue the opportunities where they have the strongest future.

Two separate worlds

One workspace serves one person, while any number of trusted collaborators and agents may work inside it as peers. The workspace holds the person's durable career model: complete experience, alternative descriptions and document variants, preferences and trade-offs, considered opportunities, evaluations, evidence and application history.

The CV and cover letter are outputs of this system, not the system itself.

CareerVector consumes the shared corpus. A CareerVector browser or desktop app may also contribute public job-ad evidence to JobCache. This participation does not merge the products or their privacy zones: private CV data, preferences, notes, evaluations and application activity do not flow into the shared corpus.

The product loop

Understand the person
        ↓
Observe the live market
        ↓
Find plausible opportunities
        ↓
Evaluate fit and explain it
        ↓
Tailor truthful applications
        ↓
Learn from corrections and outcomes
        ↺

RADAR discovers opportunities. JobCache supplies shared market memory. CareerVector adds the private person-specific interpretation. The document system turns that interpretation into action.

The intended user experience is one coherent intelligence: these jobs are unusually relevant to you; here is why; here is what remains uncertain; and here is the strongest truthful application we can construct.

AI is a resource, not the product

AI providers are changing, non-deterministic workers. CareerVector uses them to deliver the product; it must not make the product look or behave like a model picker.

Different work warrants different intelligence. Structured extraction is frequent and comparatively easy. Nuanced fit evaluation is harder. Excellent application writing needs style, truthfulness and personal context. Mechanical transformations should use deterministic code, and previously verified public facts should be reused rather than generated again.

The first-use and steady-state paths are deliberately different. A new person must receive value before being asked to wait for a multi-gigabyte model. After that first proof, CareerVector should actively move capable devices toward free local execution:

First useful action: Included AI, with no setup
    → explicit Local AI consent and background download
    → Local AI for proven task classes
    → user-provided AI for optional capacity or quality

At every stage, existing verified facts, cached results and deterministic computation precede new inference. If no zero-cost, local or authorized user-owned route is feasible, the operation waits or asks the user.

CareerVector itself does not silently buy inference. Monetary cost to the operator is a feasibility gate, not a routing preference.

User-facing AI sources

Smart Routing chooses among feasible alternatives according to the workspace's declared policy and learned evidence about quality, reliability, latency and resource pressure. Provider and model details remain available to people who want them, but they are not an onboarding requirement.

Human policy is authoritative. Learning may recommend an alternative but may not silently override a fixed choice, a hard exclusion, a privacy restriction or a spending boundary.

Included AI behaves as a resilient shared commons: credentials remain behind the CareerVector API; direct user actions outrank speculative background work; identical work is reused; smaller models handle work where they are proven sufficient; exhausted routes cool down; and the system falls back to Local AI, authorized user keys or an honest wait state. Uncertain or potentially billable operator-funded capacity is unavailable, regardless of how cheap it appears.

Local AI is represented by a persistent workspace-header status pill next to Scraping. It remains red until explicitly enabled, then exposes download, preparation, readiness and failure without blocking first use. Model weights do not download before consent. In a browser, background transfer is promised only while CareerVector remains open; interrupted downloads resume later. The full interaction contract is defined in LOCAL-AI-PILL.md.

Commercial boundary

CareerVector is permanently free. It has no premium applicant tier, paid feature, inference markup, commission or future plan to charge its users. A person may choose to pay an external AI provider through a key they own, but CareerVector is not part of that transaction and all CareerVector functionality remains free.

The sole commercial product is JobCache UI, built on the shared corpus of public job-market evidence and facts. Private workspace content, candidate access and application behavior are outside that product and are not monetized. The full decision is recorded in the business-model study.

The expected long-run improvement in model capability and price works in the product's favor: Local AI, Included AI and deterministic reuse should become more capable and cheaper over time. That progress strengthens the permanent free commitment rather than creating a later paid conversion funnel.

Collective leverage without collective exposure

More participants
    → fresher public market evidence
    → better shared facts
    → fewer repeated scrapes and AI calls
    → better routing evidence
    → lower cost and better results
    → a more useful product

The collective system may learn reusable abstractions: which routes work for which kinds of task, which public sources are reachable and which public facts are credible. It does not need to pool prompts, CVs or personal histories to do that. Normalized operational evidence and public corpus facts cross their respective learning seams; private workspace content does not.

Decisions this thesis does not settle

These are product questions, not details to leave accidentally to an optimizer. They remain explicit until evidence and deliberate decisions settle them.

1. Scarcity and fairness. Without accounts, how should Included AI divide finite free capacity between workspaces, collaborators, foreground actions, background enrichment and deliberate exploration? What prevents one actor from manufacturing workspaces and consuming the commons?
2. Quality truth. Which downstream signals honestly measure extraction, evaluation and writing quality? Hiring outcomes are sparse, delayed and confounded; model confidence and successful parsing are not correctness.
3. Privacy and provider choice. Which private fields may be sent to which external providers, what must remain local, and how should the product make this legible without turning onboarding into a compliance form?
4. Contribution and reciprocity. How transparent should browser/desktop corpus participation be, and what is the fair relationship between people who contribute scraping capacity and people who only consume shared facts?
5. Corpus governance. How are contradictory public facts, provenance, aging, correction, employer disputes and deletion requests handled while preserving an auditable market history?
6. Autonomy. How proactive should CareerVector be in discovering, judging and drafting before helpfulness becomes noise or loss of human agency?
7. Learning across people. Which routing and market lessons are safe and statistically valid to share upward, especially while the no-account model provides no strong Sybil or identity boundary?
8. Unequal devices and access. Local AI and browser contribution favor newer hardware and permissive networks. The product must not quietly make a weaker experience the price of owning a modest or restricted device.
9. Failure under real scale. How do the shared corpus, free inference supply and collaborative workspace remain responsive with thousands of actors when today's low-volume behavior stops being predictive?

The product should surface these questions through evidence and deliberate choices. Smart Routing can optimize inside the chosen contract; it cannot choose the social contract itself.