Speaker notes
Small Models, Big Ops · @marachen · 18 minutes · back to the deck
Slide 1 · Title
0:00 – 0:40
Small Models, Big Ops
Introduce yourself in one sentence, then say what the talk is not: it is not an AI talk. Set the promise — by the end they will have a six-line function they can ship on Monday.
Slide 2 · Hook
0:40 – 3:00
Your scheduling problem is not a model problem
Tell the Lakeshore story: an operations director rebuilding the waitlist by hand on a legal pad, recovering $2.1M in a year. No model. Land the line: the constraint was attention, not intelligence.
Slide 3 · Idea 1
3:00 – 6:30
The bottleneck is the follow-through
Walk through the measurement: cancelled slots were re-filled inside ninety minutes or never. Everyone in the room already knows who should get the slot. Nobody has time to make the call.
Slide 4 · Idea 2
6:30 – 9:30
Six lines beat brilliance at 2am
Contrast two deployments: a nightly batch scorer versus a heuristic on a one-minute loop. The heuristic wins because the slot expires in hours. Cadence beats accuracy when the window is short.
Slide 5 · Idea 3
9:30 – 12:30
Automate the calling, not the deciding
The trust argument. Clinics gave us the outbound texting immediately and the urgency ranking only after six months of watching. Say plainly that this was the right order.
Slide 6 · Code
12:30 – 16:00
The whole model
Read it line by line. Emphasise the two filters — can see this provider, is reachable now — because that is where the clinical safety lives, not in the sort. Then admit the sort has been unchanged since March.
Slide 7 · Takeaway
16:00 – 18:00
Ship the boring loop, then earn the model
Close with the one instruction: find the decision your team already makes correctly and automate only the acting on it. Give the handle, then stop talking.