// CASE NOTES · August 12, 2026
What shipping my own AI product taught me
Notes from building Sarana's reporting layer: where LLMs earn their keep, and where they just add noise.
By Topan Adi Pamungkas · Performance Marketing Manager · Gotrade (YC S19)
Building an AI feature for a client is one thing; running one in production for months is another. Three lessons stuck with me.
Retrieval beats memory
Every useful answer traced back to good retrieval. Tuning the index paid off more than tuning prompts.
Humans stay in the loop
Auto-send was never worth it. A one-click review step kept trust high and errors out of inboxes.
Measure or it didn’t happen
We logged acceptance rate per summary. That single metric drove every iteration.
// KEEP READING
On the same thread.
The loop I run on every account
2026 · METHOD
Migrating an MMP without losing your attribution
2026 · FIELD NOTES
What a 68% CAC drop actually took
2026 · CASE STUDIES
// ABOUT THE AUTHOR
Topan Adi Pamungkas
PERFORMANCE MARKETING MANAGER · GOTRADE (YC S19)
Performance Marketing Manager at Gotrade (YC S19), running paid acquisition for the Indonesian market as sole contributor. I started in frontend engineering in 2013 and spent three years in technical support before moving into growth, which is why I read a tracking stack as fluently as a media plan. I also build XerAds, an AI-native growth platform. I am as comfortable debugging a tracking pixel as optimising a campaign budget.
- CERTIFIED
- Apple Search Ads Certified
- Full-Stack Digital Marketing Program (RevoU)
- Advanced SEO Programme (RevoU)
- EDUCATION
- BINUS University — Bachelor of Information Systems, Computer Science
- Institut Pertanian Bogor (IPB) — Diploma, Computer Engineering
// CONTACT
A number worth digging into?
I publish what I run. If an angle is missing, say so.
Write to me