michael.wenning // ai_product
Open to AI product roles

Founder & AI Product LeaderMichael
Wenning

AI Product Management · B2B & B2C AI Products · Prompt Engineering & Evals

I build AI products end to end. Strategy to prompts to evals. Then I ship them.

Download resume↗ See the work↓ LinkedIn↗
5
AI systems in production
4×
Models orchestrated per task
93.8%
Skill-match coverage
0
Fabricated claims, by design
01 How I work

Not a slide-deck PM.
A PM with hands on the model.

Most AI PMs stop at the spec. I go all the way down: I write the prompts, build the eval suites, and own the system that ships. Here's what that looks like.

// strategy

Pick the right bets

Decide what to build and why. Tie every AI feature to a real user job, not a demo that looks good in a meeting.

// craft

Prompts & evals

Versioned prompts, real eval datasets, observability. Prompt changes get measured, not guessed. No vibes.

// ship

Own the whole path

Idea to architecture to TestFlight. I work next to engineering and stay accountable for what users actually feel.

02 Flagship
Founder & CEO · built 0 → beta

Passive

A career discovery platform for the people who aren't job hunting. You scroll, you discover, you act. No searching. I set the product strategy and built the AI systems running underneath it: parsing, matching, explanations, and interview prep.

17K+
Waitlist, sub-$1 CAC
15min
Beta filled, 10× over
+83
Net promoter score
iOS
Live on TestFlight
03 Selected systems

The AI under the hood.

Five systems I designed and shipped inside Passive. Real architectures, real tradeoffs, built to run in production.

// orchestrated across 3 model providers, chosen per task
GPT-4.1 OpenAI resume tailoring Perplexity live web research Claude Sonnet Anthropic coaching Claude Opus Anthropic synthesis

Grounded Match Explanation

// why this job, in words you can trust
architecture
Problem LLM explanations drift from the facts. Hiring is too high-stakes to let a model invent a reason you matched a role. How I built it A deterministic evidence object feeds two tiers: an instant template answer, then an async LLM-written version. Everything passes a faithfulness gate before a user sees it. The core is deliberately non-RAG so the model can't wander off the evidence.
in evidence object tier 1 template · instant tier 2 LLM · async gate faithfulness out explanation
Result Explanations that are fast, grounded, and never make something up.

Resume Tailoring Engine

// tailors a resume to a role, honestly
v3 · live
What it does Tailors a resume to a specific job description and scores it before and after across five enterprise ATS platforms: Greenhouse, SmartRecruiters, Workday, Lever, iCIMS. Returns structured JSON plus a full changelog. Why it matters Most AI resume tools inflate. They invent credentials to juice a match score. This one can't. An absolute no-fabrication rule with worked examples sits at the top of the prompt, every change is logged with what changed, why, and which requirement it maps to, and the resume grades itself before and after. The lift is measured, not claimed.
score before, vs JD parse required · preferred · implicit optimize integrity-bounded score after log changelog
no laundering The rule is concrete: 'Assisted with' can't become 'Led.' 'Contributed to' can't become 'Owned.' If a tool or certification isn't in the original resume, it never gets added. An empty field beats a plausible-sounding lie.
// stack GPT-4.1 · temp 0.2Before/after ATS scoring5 ATS platformsAnti-fabrication ruleAuditable changelogStructured JSON

C.R.A.F.T. Interview Intelligence

// turns a job description into a real prep plan
v2.0
What it does A four-phase engine that turns one job description into a full prep plan. It researches the company and role live, builds a behavioral story bank from the candidate's own history, coaches on presence and delivery, then writes the sharp questions they should ask the interviewer back. It ships as a formatted PDF. Why it matters Interview prep is the part of job searching people dread and do worst. Generic advice doesn't move the needle. This is specific to one candidate and one role, and it lands in minutes. The hard part wasn't any single prompt. It was making the four phases compound, so each one writes structured context the next phase reads.
phase 1 research · Perplexity phase 2 story bank phase 3 presence prep phase 4 reverse Qs · PDF
state compounds The company's values, captured word-for-word in phase one, come back in the phase-four questions. Not 'how important is transparency?' but 'Radical transparency is one of your stated values, give me a recent example.'
// stack Perplexity researchClaude Sonnet + Opusn8n pipelineStructured outputFormatted PDFTested on real candidates

Resume Parser + Skills Taxonomy

// the data layer matching runs on
v3.1
What it does The parser reads a resume and returns work history, seniority, and a flat list of skills. That list flows into a canonical skills taxonomy I designed from scratch across ten-plus white-collar functions, with a 17-field schema for every skill. Why it matters Matching and resume tailoring are only as good as the skills data underneath them. Job titles lie. Resumes use a hundred names for the same skill. Most systems choke on that. I treated the taxonomy as a product: an alias graph does all the runtime normalization, so the parser stays dumb and fast while the graph maps 'GA4', 'Google Analytics', and 'web analytics' to one canonical skill. Titles do double duty too, inferring skills a candidate never spelled out.
61.6% → 93.8% Skill-match coverage after I rebuilt the alias layer bare-form first. The lesson: build umbrella canonicals and bare-form aliases from day one, or the parser silently drops a third of real skills.
// stack Ontology design17-field schemaAlias graphCosine-similarity matchingTitle-based inferenceATS ingestion · Greenhouse, BambooHR

LLM Observability & Evals

// prompts you can change without fear
Langfuse
What it does Prompt versioning decoupled from code deploys, with datasets and eval runs wired in. Product owns the prompts. Engineering owns the code. Neither blocks the other. Why it matters This is how a small team ships AI fast without breaking it. Every prompt change is measured against a dataset before it goes live.
04 What I work with

The toolkit.

01Prompt engineering 02Eval design 03Multi-model orchestration 04LLM observability (Langfuse) 05RAG 06Structured outputs / JSON 07Faithfulness & grounding 08Taxonomy & ontology design 09Workflow automation (n8n) 10Product strategy 110 → 1 shipping 12iOS beta programs 13ATS integrations
05 Background

I know hiring from the inside.

Before AI, I spent 15 years in recruiting and staffing. I founded two companies in the space and saw every broken part of how people find work.

That's the whole reason I build for it now. I'm not an AI person who picked hiring as a market. I'm a hiring person who learned to build AI, so the products solve problems I've actually lived.

Based in Zagreb, Croatia. Building Passive.

2024 → now

Founder & CEO, Passive

AI career discovery platform. Product strategy plus the AI systems underneath.
earlier

Founder, Wenn

Wenn.AI, a B2B AI SaaS.
earlier

Founder, Hire Metrics

Recruiting and staffing.
15 yrs

Recruiting & staffing

The foundation for everything I build today.
06 Contact

Building in AI?
Let's talk.

Email me↗ LinkedIn↗ Download resume↗