Career case-study system

An agentic system that runs a job search,
built so its reliability is the proof.

It sources roles, tailors applications, and ships them through a human-approval gate. How it's engineered is the demonstration: the guardrails, the honesty, the oversight an operator brings to AI operations and solutions engineering.

Human-in-the-loop Multi-skill orchestration Honest-by-construction Deployed & operated
Read the architecture

The spine

A human and an AI, building in the open.

This system wasn’t commissioned and handed off. It was designed turn by turn: a human setting direction and drawing the hard lines, an AI doing the building, and a written decision record holding every fork accountable.

The interesting part isn’t that an AI wrote code. It’s the operating discipline around it: explicit invariants that survive across sessions, a guardrail that no automation may cross, and a rule that nothing ships unless it’s true. The collaboration is the artifact, and the way it’s run is the evidence.

Layer B · System deep-dive

The architecture

Three pillars. An engine of composable skills, a loop into reality that actually applies for jobs behind a human gate, and this showcase, the projection that makes the system legible to someone who wasn’t in the room.

Pillar 1

Engine

A set of single-purpose skills (discover, tailor, prep, assemble) that compose into a pipeline. Each is a black box with a clean input and output; the orchestration is the design.

Pillar 2

Loop into reality

The engine doesn’t stop at “generated”. It drives a real funnel, sourced to closed, through a human-approval gate, then captures the outcome. Real applications, real results.

Pillar 3

Showcase

This site. A statically-generated public projection: curated content and sanitized outcomes only, walled off from the private substrate. Designed, built, and deployed as part of the artifact.

The orchestration

How an application moves

From a posting in the wild to a tracked outcome: every stage named, the human gate in the middle, nothing hidden except the parts that are the moat.

  1. Discover career-source

    Finds real, current postings and fit-scores each against the candidate's history. It curates and ranks; it never maximizes count.

    sourced
  2. Tailor career-jd-match · career-resume

    Decomposes the JD, scores fit, names the honest gaps, and produces a tailored résumé and pitch. Every claim traces to a real accomplishment.

    tailored
  3. Assemble career-apply

    Builds the submission package, a tailored résumé plus a cover letter, and presents it for review.

    awaiting-approval
Human approval gate

No application leaves without an explicit human yes. There is no auto-approve path. The gate is the invariant; its transport (a phone tap, a local prompt) is swappable.

Sourced
Tailored
Awaiting approval
Responded
Interview
Closed

A daily scheduled routine watches live applications and nudges for the next decision. It only detects; it never advances an outcome on its own. Outcomes derive from each application's furthest-reached stage, so the conversion story stays honest even after an application closes.

The guardrail philosophy

The lines that don’t move

A system you’d trust to act on your behalf is defined by what it refuses to do. These invariants survive every change.

  1. 01

    Human approval before any send

    The system does about 95% of the work and stops at the boundary that matters. Every outbound application waits for an explicit human yes. Appropriate autonomy with human oversight is the exact maturity these roles screen for.

  2. 02

    Official channels only

    No form-stuffing against sites whose terms forbid it; no auto-apply. Submission is a pluggable transport with a hard terms-of-service guard, and it degrades to a human send whenever a channel requires it.

  3. 03

    Truthful by construction

    Every claim, on this site and in every application, traces to a real role or venture note. Metrics are never fabricated. With no submitted applications yet, this page shows the system and the approach, not invented numbers.

  4. 04

    A hard privacy boundary

    This public site never imports the private substrate. Only curated content and aggregate, sanitized metrics ever cross over, by a manual reviewed copy-over, not an automated sync. No PII; no employer identity without consent.

  5. 05

    Architecture, not implementation

    What you see is the orchestration approach, the guardrail philosophy, and the measured outcomes. The elicitation technique, scoring rubrics, and engine code stay private. Here is what was designed and what it produced, not the source.

  6. 06

    Quality over volume

    The system curates targets; it never maximizes application count. The story this funnel tells is conversion, not top-of-funnel width.

Live outcomes

The numbers, when they’re real

Applications sent
Response rate
Interviews

This panel is wired for a sanitized, aggregate metrics feed: counts only, never an employer identity or any PII. It’s blank because no applications have been submitted yet. When the funnel runs, real numbers appear here; until then, showing nothing is the honest choice.

The system, in the wild

What it produces

Every candidate the system runs gets their own deep-dive page: generated from their real history, customized to them, deployed to its own home. Here is what that looks like.