Sean Acres

AI product leader who managesbuilds

Agent harnesses, generative pipelines, and the evals that decide whether the behavior is good enough to ship.

Founding PM to VP of Product across seven 0-to-1 launches: a #1 iOS game, a title on a 180M-subscriber franchise, a 3.6M-MAU streaming org. Owned P&L at three companies; hired and led product, design, and growth orgs across the US and India, from seed through Series C.

Currently Senior Product Manager at Weekend, improving the agent harness the studio builds games with, version after version. Its output went from months per title to five launches in five weeks. Working in voice AI since 2017. I ship the Python, TypeScript and Swift, and the evals behind them.

Built on my own time

Four products. One is live, one is a waitlist, two are private repos, and the page says which is which. None of them has real external users, so nothing here claims any.

Autonomous car-buying agent

Raretta

Name a car and a ceiling. It hunts dealer inventory around the clock and negotiates by email while you do nothing.

Pulls live inventory from manufacturer and used-car APIs, so the ranking sees the real market rather than one dealer group's lot
Scores every listing on what it would cost delivered, not the sticker
Blocks its own outbound email whenever the model invents a price, sending nothing rather than something generic

raretta.com · signup closed by construction

Live

On-device AI assistant for iPhone

Tellspike

Ask it anything without the question leaving your phone. When it does reach the cloud, it sends an anonymized rewrite.

Every cloud message carries a receipt showing the exact text that left the device, so the privacy claim is checkable rather than asserted
The free tier's silence is enforced by a test that fails the suite on any outbound request, analytics included
Privacy priced as the product: free on the device, metered in the cloud

tellspike.com · app in development, not released

Pre-launch

Asset compiler for game developers

Tosskeep

Describe any game asset and get engine-ready output, with an AI quality gate deciding what is good enough to keep.

Meta-evaluation grades the judges themselves against human approve and reject verdicts
Measured that post-processing closes a 28-point quality gap, letting a $0.03 model match a $0.36 purpose-trained one
Its architecture became the art pipeline at work, rebuilt there in two weeks

private repo · no public build

Private

AI character SDK

animachord

Characters remember you, decide whether they trust you, and gossip about each other behind your back.

Grounded in published psychology: Big Five, attachment theory, Tulving's memory taxonomy, Gottman ratios, Ebbinghaus decay
The relationship engine is LLM-free: the part deciding whether two characters drift apart has no model call in it
Reimplemented in Swift rather than bound to, with CI failing the build if the two runtimes diverge on disk

private SDK · gated demo in progress

Private

Case studies

Four decisions, the evidence each one rests on, and what I would do differently.

The claim ladder

Every level of privacy evidence mapped to the exact marketing copy it entitles, including the line we had to retract when the measurement came back weaker than the claim.

Tellspike & animachord

Measuring a claim instead of asserting it

An n=100 paired eval against three attacker tiers. The useful result was the one that failed: the category the rewrite could not protect, and why that was load-bearing.

Tellspike

Cost per approved asset

A cheaper model that fails review twice is the expensive one. Ranking generative models on the outcome that matters changed which model we would have bought.

Tosskeep

On-device by default, as a pricing model

Privacy usually lives in a policy page. Here it is the architecture and the price list: free and unlimited on the device, metered in the cloud, falling back to the device at zero credits.

Tellspike