The day in six minutes

6 chapters

In this episode

1. 0:00Welcome
2. 0:37The opening keynote

   Rhea Kulkarni

   Read the session notes
3. 1:35AI agents in production

   Meera Iyer, Arjun Bhatt, Sana Qureshi, Tomás Reyes

   Read the session notes
4. 3:06Hiring for AI product teams

   Farah Siddiqui, Kabir Shah

   Read the session notes
5. 4:20Five things we shipped

   Neha Joshi, Rahul Menon, Aisha Khan, Rohan Das, Leela Rao

   Read the session notes
6. 5:23Three things to take home

1. 0:00MayaHey, welcome to the Day 1 recap from the TPF UnConference in Bengaluru. I'm Maya.
2. 0:08DevAnd I'm Dev. If you couldn't make it today, or you were stuck in one room while the good stuff happened in another, this is the whole day in about six minutes.
3. 0:23MayaFour sessions, and one big theme. Everyone was talking about agents, but almost nobody was talking about the model. They were talking about trust.
4. 0:34DevLet's start at the top.
5. 0:37MayaRhea Kulkarni, who runs product at Northwind Health, opened the day. Her pitch was simple: the room sets the agenda, not the organisers.
6. 0:49DevAnd she was blunt about what to bring. Her line was:
7. 0:54Dev“If you've already solved it, write a blog post. Bring the thing that's keeping you up at night.”

   Rhea Kulkarni, quoted from the session. Find it in the notes

8. 1:03MayaThe part that stuck with me was the roadmap story. Northwind threw out quarterly roadmaps and moved to six-week bets, one metric each.
9. 1:14DevAnd they shipped half as many things. Retention went up. She said that's not a coincidence.
10. 1:22MayaShe also said agents now write the first draft of every spec at Northwind. But a named PM still signs the bottom of it. Ownership doesn't move.
11. 1:35DevWhich set up the morning panel nicely. Meera Iyer from Globex Payments moderated three people running agents with real customers.
12. 1:45MayaAnd they all agreed: the hard part isn't the model. It's knowing when the agent is wrong.
13. 1:53DevArjun Bhatt from Initech runs four hundred real cases on every prompt change. If the pass rate drops, the release is blocked. Same rule as tests.
14. 2:06MayaSana Qureshi from Hooli Labs pushed back, which I loved. She had eleven customers, so her eval set was, in her words, me, a spreadsheet and a lot of coffee.
15. 2:20DevAnd Arjun's answer was that it counts. Fifty good cases beat a fancy tool with no cases.
16. 2:29MayaThe surprise came from Tomás Reyes at Umbrella Retail. They pushed accuracy from eighty-one to eighty-nine percent, and nobody noticed.
17. 2:38DevWhat people did notice was when the agent started showing its steps. As he put it:
18. 2:46Dev“Our agent got more trusted when we made it slower and showed every step. People want to see it think.”

    Tomás Reyes, quoted from the session. Find it in the notes

19. 2:56MayaPricing stayed messy. Two panellists moved off per-seat pricing, and one moved back. Nobody has this figured out yet.
20. 3:06DevAfter lunch, Farah Siddiqui, who heads talent at Globex, sat down with Kabir Shah from Contoso to talk about hiring.
21. 3:15MayaGlobex killed the whiteboard case study. Candidates now get a laptop and twenty real model outputs, and they decide which ones they'd ship.
22. 3:26DevWhat she's watching for is whether you slow down. The good candidates get suspicious of the outputs that look fine.
23. 3:36MayaShe had a great line on experience, too:
24. 3:41Maya“Ten years of experience can be one year, ten times. I'd rather see what you built last month.”

    Farah Siddiqui, quoted from the session. Find it in the notes

25. 3:50DevAnd on job titles, she said there's no such thing as an AI PM. There are PMs, and some of them have spent real time with the models.
26. 4:03MayaThe audience question I keep thinking about was: if agents do the junior work, how do juniors learn? Farah said it's what she loses sleep over. For now, juniors review the agent's work before seniors do.
27. 4:20DevThen five lightning talks, seven minutes each, all about things people actually shipped.
28. 4:27MayaNeha Joshi's team replaced a fourteen-field signup form with a ninety-second voice chat. Completion went from forty-one to sixty-eight percent.
29. 4:36DevWhich, with a voice track at Builders Day tomorrow, is a pretty good hint.
30. 4:43MayaRahul Menon switched off Umbrella's do-everything chatbot and shipped three single-purpose buttons instead. Support tickets dropped.
31. 4:51DevAisha Khan says you don't need an eval tool to start. You need fifty examples in a spreadsheet and a Friday afternoon.
32. 5:02MayaRohan Das charges per finished report, not per seat. His line: nobody wants to buy seats for a robot.
33. 5:11DevAnd Leela Rao closed with a warning. AI features dip around week four, when the novelty runs out. Have the second reason ready.
34. 5:23MayaSo, if you take three things home from Day 1.
35. 5:28DevOne, start narrow. Pick one job the agent can finish end to end.
36. 5:35MayaTwo, write your evals early, even if it's fifty rows in a spreadsheet.
37. 5:41DevAnd three, show the steps. People trust what they can watch.
38. 5:47MayaThe full notes, quotes and transcript excerpts for every session are on the platform. Tomorrow is Builders Day, so expect a lot more demos.
39. 5:59DevSee you then.