I designed and shipped Smart Resume AI — a Chrome extension that turns any job posting into a tailored, truthful application. One fabricated number can decide whether a candidate gets the interview, so the system is built so the model can't invent: every output grounded, diffed, and human-reviewed.
A product case study in making AI trustworthy where it has never been — built by a solo product lead directing an AI engineering fleet.
Watch the 3-minute product tourGeneric generators invent metrics and produce fluff that ATS semantic filters reject. In a document that must be 100% truthful, a 95%-accurate model is unusable.
A verified Experience Vault feeds a 5-stage pipeline. Prompts carry zero-tolerance
metric rules; unknown companies return companyInfoAvailable: false instead
of fiction; every output renders as a color-coded diff for human review.
51 in-browser diagnostics + 302 CLI assertions, including dedicated anti-hallucination, unavailable-company, and source-priority suites — all passing.
Three narrated minutes, end to end: the posting goes in, the vault and guardrails do their work, and what comes out is a diff you audit — not text you're asked to trust.
Anyone can call an LLM API and get resume-shaped text back. The product question that mattered: how do you ship AI into a workflow where a single fabricated number — an invented "$4M ARR" on a resume — is a disqualifying, unrecoverable failure?
“Mine what I actually did, in my own verified words. Research what the company actually is. Tailor the framing to what this role actually needs — and invent nothing.”
That constraint set the architecture: a grounded evidence store, a staged pipeline, hard output contracts, and a human sign-off on every diff.
The extension captures the job description straight off the ATS page (Ashby, Greenhouse, Lever, Workday — including shadow-DOM traversal), then runs staged generation where each LLM call does one bounded job. Deterministic code does what code does best.
Per-ATS DOM parsers scrape title, company & full JD in place — no copy-paste.
Deconstructs responsibilities, mandatory stack, seniority, company type.
Maps vault stories against requirements; returns relevance-ranked matrix + per-dimension fit scores.
Live web snippets → optional LLM synthesis. Checklist-gated: unchecked topics never enter the prompt.
Tailored bullets aligned to JD pain points — metrics locked verbatim.
Strategic hooks from research, proof points from the vault — run in parallel with Stage 4.
Ashby, Greenhouse, Lever, Workday, or any careers page — the extension reads the posting in place. Nothing to copy-paste.
One click scrapes the JD and the 5-stage pipeline runs with a live stepper.
Tailored resume + cover letter, plus a <200-word recruiter pitch and hiring-manager DM.
Live split editor with a vector PDF preview — then auto-attach it to the application form.
Each subsystem answers one failure mode of naive generation. Open a tab, then “Inspect the engineering” for the actual prompts, formulas, and fallback logic — this is the layer a product lead actually designs.
The #1 reason AI resume tools fabricate is that the model has no source of truth. This product's moat is a structured evidence store the LLM is forced to cite.
The JSON schema itself is assembled from active toggles — the model can't wander into disabled topics, and users on trivial postings don't pay for analysis they don't need.
| Dimension | Formula | Score | Band |
|---|
A significant share of the early effort went into stress-testing the hypothesis itself — was truthful AI tailoring viable, and what exactly should be generated? — before committing to product code. The build then ran as parallel flows in isolated worktrees, with the human setting direction, reviewing every diff, and making the kill decisions. A working preview of AI-era product team topology.
Before building, the core question was pressure-tested from multiple angles: can an LLM tailor a resume truthfully, and what exactly should it produce? Exploratory flows proved out the generation approach before any of it became product.
Viability FirstCandidate stacks were explored simultaneously in separate worktrees. The wrong architecture — a bundler that fought the extension platform — was killed while it was still cheap to kill, and the product was rebuilt on fundamentals that held.
Kill Decisions EarlyATS capture, the staged pipeline, vector PDF export, and one-click form upload — the end-to-end loop that makes the product usable on a real job posting.
End-to-End LoopExperience Vault, the diff engine, fit scorecards, and zero-hallucination contracts — the layer that made AI output reviewable and trustworthy, plus the discipline to cut shipped features nobody loved.
Trust & Scope DisciplineAI expands capacity; the human owns the judgment. The same principle that governs the product's human-in-the-loop review governs the team that built it.
The anti-hallucination policy is enforced in prompts and verified in code: dedicated test suites assert that metric-preservation rules exist in every generation prompt and that unknown companies degrade honestly. All green at time of writing.
companyInfoAvailable: false and say so — no invented HQ, funding, or headcount.Distilled from shipping this system and steering the fleet that built it.
The verbatim-metric rule blocks do more for product trust than any UI polish. Treat prompts as versioned interfaces: tested for regression, explicit about failure modes, and boring on purpose. “Invent nothing” is a spec, not a vibe.
The system automates scraping, mining, drafting, and formatting — then stops. The diff engine, revert controls, and export gate exist so a human makes the final call on every claim that carries their name. Same rule worked for running the AI fleet.
When the system doesn't know the company, it says so. When a story has no metrics, the bullet stays qualitative. “I don't know” rendered in the UI is a feature — it's what makes the confident answers believable.
I'm looking for product leadership roles where the interesting question isn't “where do we bolt on an LLM?” but “how do we make model output safe, grounded, and accountable enough to ship?”
This case study is one instance of a pattern I keep reaching for: technically hard systems, honest constraints, human judgment preserved where it matters — whether that's financial infrastructure, developer platforms, or applied AI.