Benjamin Jacobs builds AI agents and automations that ship in days.

Building

Every Rep, my PayPal hackathon entry, now in judging. Taking contract work for AI agents and automations.

Shipped

Contact

[email protected]
Honolulu, HI
Contract or full-time
US hours

80 seconds: real systems, recorded running. Sound on.
01

Every Rep

Friends pledge money for each push-up, a camera counts the reps, and people pay only for the reps that happened.

Problem
Pledge-per-rep fundraisers charge a flat amount up front or chase people later. Nobody wants to pay for reps that never happened.
Built
PayPal holds each donor's limit. The camera counts reps live, then charges reps times rate when time is up and voids the rest. The money goes to the cause. Built for the PayPal AI hackathon.
Stack
PayPal Orders (authorize, capture, void) and Payouts, in-browser pose tracking, Node. Test mode, no real money moves.
The 56-second hackathon demo. Live at everyrep.swellstackco.com (test mode).
02

Lantern

Three short daily voice check-ins that catch early memory and speech change, with a one-page summary for the doctor.

Problem
Early memory and speech change is easy to miss between doctor visits, and families notice too late.
Built
Answers are scored against the person's own 90-day baseline. A drift detector flags a sustained change calmly to family, and a study dashboard tracks a pilot group and exports de-identified FHIR and CSV. 13-tool MCP server, consent flow, 98 automated tests.
Stack
Node, Express, SQLite, MCP (Streamable HTTP), EWMA drift detection, FHIR export. Code on GitHub ↗
Hackathon build. Every flag is labeled "a reason to talk, not a diagnosis."
03

A salon website with a concierge on it

Live, and the owner changes things herself without calling me.

Problem
A busy salon with a dated site, and staff answering the same booking and service questions all day.
Built
A new site, an AI concierge named Ellie that answers service questions and points people to the right stylist, and a self-serve admin panel.
Stack
Custom site, AI chat concierge, admin panel.
Salon website with the Ellie chat concierge open on the right, greeting a visitor
Ellie on the live site.
04

Footage in, 90 posts out

90 posts published July to October, each one checked live after it went up.

Problem
One Hawaii real-estate firm, four platforms, raw phone footage, and nobody with time to edit, caption and post it.
Built
A pipeline that turns footage into finished reels, lints captions for compliance, publishes to four platforms, then reads each post back to confirm it went up. Plus a dashboard where the client approves.
Stack
ffmpeg, Whisper, Cloudflare R2, a client review dashboard.
SwellDeck client dashboard showing work completed, platforms connected and an approval queue
The review dashboard, with demo data.
05

An AI receptionist that answers the phone

About 500 ms to the first spoken word. Built and parked while I focus on other work.

Problem
Small businesses miss calls, and a voice bot that pauses for two seconds sounds broken.
Built
Audio streams in over Twilio, gets transcribed, goes to an LLM, and the reply is spoken back. Stripe billing and a 133-page SEO site came with it.
Stack
Twilio Media Streams, Deepgram, Groq, Cartesia, Stripe.
A 45-second mock call: Mia books a roofing estimate.
06

A product assistant that has to be right

Live, and every release has to pass the accuracy test set before it goes out.

Problem
Architects ask manufacturers which product fits a job. A wrong answer is worse than no answer.
Built
A white-label, multi-tenant assistant for building-products manufacturers. I have also designed a second phase that reads plan sets, picks products by rule and sends low-confidence picks to a person.
Stack
Multi-tenant LLM assistant, eval gate in the deploy path.

How I work

  1. 1
    Two-week paid pilot

    You pay for two weeks of real work. If it is not useful, we stop.

  2. 2
    Three named deliverables

    We agree on three things I will hand over, written down before day one.

  3. 3
    A demo every week

    You see it running on a call, not in a slide deck.

Fifteen minutes is enough to tell if it is a fit.

Best fit: you have a manual process or an LLM idea and nobody to build it. Not a fit: pure research roles, or anything where I would never touch production.