LLM product manager for startups

Startups face a specific LLM trap: the technology makes impressive features cheap to demo and expensive to make dependable, and a startup's credibility can't absorb the gap. An enterprise survives a mediocre AI launch; a twelve-person company betting its next round on an LLM product can't. What that company needs is product management that moves at startup speed while refusing to ship on vibes.

Founder speed, production standards

I'm a two-time founder who built on LLMs daily at WisOwl AI (since acquired): agentic recruiter systems and semantic matching on FAISS and Supabase pgvector, which served 10,000+ jobseekers acquired with zero paid marketing. My operating principle, learned across both my companies, is that speed is a feature, and the fastest sustainable loop is ship-measure-iterate against real behavior. Applied to LLM products, that means small scoped bets with eval harnesses from day one. No six-month AI platforms, and no YOLO-shipping whatever demoed well on Friday.

How I run LLM product at a startup

• Pick features where variance is tolerable. First LLM features should live where a mediocre output disappoints mildly rather than damages: drafting, summarizing, matching with human confirmation. Trust compounds from there.
• API-first, always, at this stage. Model APIs plus smart prompting and retrieval win on speed and iteration cost. The fine-tuning conversation can wait until an eval proves the ceiling.
• Evals sized to reality. A startup doesn't need an evaluation platform; it needs fifty golden cases per feature and the discipline to run them before each release. I set that up in days.
• Unit economics on the PRD. Token cost per user-action, projected at success-case volume, decides pricing and caching strategy before launch, because a feature can succeed with users and still wipe out the margin.

I've also spent eight years at CaaStle running growth product on a ~$50M+ ARR portfolio, so when the LLM feature works, I know how to wire it into activation, retention, and pricing. Engagements: fractional (1–3 days a week), or a single feature taken from scoping to launch as a fixed project.

Frequently asked questions

Should our startup build an LLM feature or is it hype?

Run the boring test: does the feature reduce real user effort on a frequent task, and can you tolerate its worst output? If yes twice, it's a candidate. If it exists so the deck can say "AI," skip it. Users notice, and so do investors at diligence.

Which model should we build on?

Start with a top-tier API model to prove value, then route down-tier where evals show cheaper models suffice. Model choice is a routing decision that changes monthly; the eval suite is the asset that makes it changeable safely.

How much should an LLM MVP cost to build?

Weeks, if scoped honestly. The build is rarely the bottleneck. The costs that surprise startups are eval time and iteration cycles after contact with real inputs.

Can you work with our tiny team — two engineers and a founder?

That's my natural habitat. At that size I'm hands-on: writing the golden datasets, prototyping prompts, and pairing on the integration.

Related pages