Day 1 · Sat 21 Nov
Lightning talks: five things we shipped this year
PreviewMade-up talks and speakers until the real recordings come in.
In two sentences
Five builders, seven minutes each. The common thread was cutting scope: a voice chat that replaced a form, a chatbot that got switched off, and evals that cost less than a team lunch.
010:00
Neha Joshi: voice onboarding in 30 days
Replacing a long signup form with a short voice conversation, and what they had to fix for accents and noisy rooms.
Neha Joshi from Initech opened with a signup form that had fourteen fields, and a team that had spent a year A/B testing the button colour. They replaced it with a ninety-second voice conversation that asks about the customer's business and fills the form in the background.
Completion went from 41% to 68%. The hard part wasn't the model but the real world: accents, noisy shop floors and people who answer three questions at once. Her takeaway was that nobody wants to fill a form, but everybody will tell you about their business if you just ask.
“Nobody wants to fill a form. Everybody will tell you about their business if you just ask.”
027:30
Rahul Menon: killing our chatbot
Why a do-everything chatbot lost to three buttons that each do one thing well.
Rahul Menon's talk was about switching something off. Umbrella Retail had a chatbot that could track an order, change an address and tell a joke, and customers used it for one thing: asking for a human.
They replaced it with three single-purpose buttons (track my order, change delivery, talk to someone) and support tickets dropped. He called the off switch the best AI feature they shipped all year.
“The best AI feature we shipped this year was the off switch on our chatbot.”
0315:10
Aisha Khan: evals on a budget
Fifty labelled examples, one spreadsheet, a weekly review. Enough to catch most regressions.
Aisha Khan from Hooli Labs said everyone asks her which eval tool to buy. Her answer is a spreadsheet and a Friday afternoon. Her team keeps fifty hand-labelled examples and reviews them weekly, which catches most regressions before customers do.
Her point was to start small and start this week: fifty examples in a spreadsheet beat zero examples in a fancy tool.
“Fifty examples in a spreadsheet beat zero examples in a fancy tool.”
0422:45
Rohan Das: pricing an AI feature
Moving from per-seat to per-report pricing, and the one customer it didn't work for.
Rohan Das, founder of Acme Reports, moved from per-seat pricing to charging per finished report. He said it made every sales call easier, because customers could connect the price to something they actually wanted. Nobody wants to buy seats for a robot; they want to buy the report.
It didn't work for every customer. One large account wanted a predictable annual bill, so Rohan offers a prepaid bundle of reports for them.
“Nobody wants to buy seats for a robot. They want to buy the report.”
0530:20
Leela Rao: retention after the novelty
The week-four dip in AI feature usage, and the nudges that brought people back.
Leela Rao from Northwind Health closed with a pattern she'd seen on every AI feature they launched: a huge first week, steady for three weeks, then a cliff around week four when the novelty runs out.
The features that recovered had a second reason to come back ready before launch, like a weekly summary or a reminder tied to real work. Her advice: plan the second reason before you ship the first.
“Week four is where the novelty runs out. Have the second reason ready.”
How this was made
Transcribed by 
The full 38-minute recording was transcribed with ElevenLabs Speech to Text, with each voice matched to a speaker. This write-up was drafted from that transcript, then checked by the TPF team. It’s a detailed summary, not a word-for-word record. Quotes are exact and timed to the recording.
- Recording
- 38m
- Speakers
- 5
- Quotes
- 5
Transcribed by![]()
