Playbook

Updated

AI Product Playbook.

How I approach building AI products people can trust — workflows before models, humans in the loop, and shipping that survives Monday morning.

What
Central AI product philosophy
Who
Founders & builders adding AI to real work
Why
Models are cheap; trust is not
Next
How I build · Patterns
01

Finding the right problem

Start where someone already compensates — spreadsheets, memory, inboxes. If nobody is struggling, there is nothing useful to automate.

AI is not a reason to build. Friction is.

Ask: what fails under time pressure? What creates professional risk when wrong? Those problems deserve systems; novelty does not.

02

Designing workflows before models

Map the path humans already repeat. Place AI as a stage — extract, draft, rank — never as the owner of the system of record.

OCR without a review queue is not document intelligence. Chat without a matter file is entertainment. Design the workflow first; choose models second.

Document Intelligence pattern →

03

Human-in-the-loop systems

The review screen is the product. Confidence beside every field. Accept, edit, reject. No silent merges into client matters.

Experts need acceleration, not abdication. That constraint shaped NepalIPMS OCR intake and every AI feature I will ship into a system of record.

HITL decision →

04

Evaluating AI features

Demo metrics lie. Ask whether the feature reduces rework, shortens a desk task, or increases trust on a live call.

Use the AI evaluation checklist. If you cannot name the review surface and the failure mode, do not ship.

05

Shipping safely

Expand-contract migrations. Feature paths you can disable. Smoke the desk-critical path with real-shaped data. Prefer edge stacks a small team can operate.

Safe shipping is boring on purpose — renewal season does not care about your launch tweet.

Release checklist →

06

Measuring usefulness

Track outcomes desks feel: time to complete intake, rate of human edits, search trust (retries, abandoned queries), missed deadlines prevented.

Ignore vanity user counts. Prefer evidence you would put on /evidence.

07

Common mistakes

Fuzzy-first search that looks smart and breaks trust. Treating OCR as understanding. Auto-writing the system of record. Building AI as the product instead of a stage. Shipping without a review UI. Optimizing demos over Monday mornings.

I have made several of these. The lessons are public.

Lessons from production → · What changed my mind → · Principles →