The problem
Short-form content is how creators and small businesses get found, and producing it is a grind: pick the best moments, cut vertical clips, write captions that sound like you, post to every platform at the right times, and confirm it all actually went out. Most people either burn hours doing it by hand or pay an editing service to do it generically. I wanted the third option: a machine that does the production work and keeps the person's voice.
What I built
PeakClips is a multi-tenant SaaS platform that automates that pipeline end to end. For streamers, it ingests Twitch clips, scores them for potential, runs them through a safety classifier, reformats them to vertical with face tracking, writes captions in the creator's own voice, and schedules publishing across nine platforms, from TikTok and Instagram through LinkedIn and Bluesky. A reconciliation job then confirms each post actually published. The same backbone powers an image-first pipeline for other verticals, including a version for real estate agents.
The part I care most about is the voice system. Each account gets a versioned voice profile extracted from things the creator actually wrote, with a guard that refuses to train on AI-generated text, so the profile can't drift into generic model-speak over time. AI here is production infrastructure with budgets and audit trails: a central model registry routes each job to the cheapest model that can do it well, every call logs its token costs per tenant, and image generation checks a monthly budget before spending a cent.
I also did all the marketing a founder can't skip: the positioning, the landing page and onboarding flows, the legal pages, and a 30-post SEO blog aimed at comparison and alternative searches.
What it does now
The platform runs in production daily for operator-owned accounts. Sun Coast SUP's automated morning report publishes through it every day, and a build-in-public image pipeline runs on the same rails. The creator and realtor products are built through onboarding, checkout, and publishing, and are held at the gate deliberately: my own end-to-end audit concluded no real creator should be invited or charged until a full smoke test passes from ingestion to confirmed publication. It hasn't yet, so nobody has been billed.
What went wrong
That launch gate is the honest part of this story. The most recent full-pipeline test stalled when Twitch wouldn't resolve playable media for the test clip, which is exactly the kind of failure you want to find before a paying customer does. Earlier, a caption bug at the publishing vendor posted only the first line of multi-line captions, and their own documentation bot gave me confidently wrong advice about the fix. Building on third-party platforms means their failure modes become your reliability problem, and I underestimated how much of the work would be hardening rather than features.
What it demonstrates
Product marketing and systems engineering as one job. I can write the positioning, then build the pipeline that delivers on it, and I know where AI belongs in that pipeline: metered, guarded, and doing production work.