I learn by building, measuring and writing it down.
Courses help, but the lessons that stick come from shipping something real and checking it honestly.
Learning now
Agents and evaluations with Claude. Project 1 taught me retrieval, citations and strict evals. Project 2 is teaching me tool use, escalation and approval flows.
Lessons I learned the hard way
A generous metric hides real misses.
The note: my retrieval check passed on luck until I made it strict, and the honest score was lower.
Read the note →Measure the way users actually type.
Testing the questions visitors type first took first-question success from 43% to 100% on the hotel assistant.
See the results →Cheaper can be just as good, if you measure it.
On the same 54 questions, Haiku 5.5 matched Haiku 4.5 within run-to-run range at about 5× lower cost per 1,000 questions.
See the comparison →AI is one layer, not the whole system.
Most of what makes an AI product work is ordinary good engineering.
What I believe →
Education
Master of Computer Applications (MCA) and Bachelor of Science (B.Sc.), Acharya Nagarjuna University, India
The long road
2014 Java → 2023 production AI → nowThe full story →