rajatSingapore ·

Writing

Notes from building applied AI systems, product loops, data pipelines, and the small engineering lessons worth keeping. Recent notes show what is fresh; shelves collect longer trails; the archive stays quiet until search or tags make it useful.

Most recent

30 MarThompson Sampling: how explore meets exploitA Bayesian bandit algorithm that solves the exploration-exploitation tradeoff. How it works, why it's elegant, and where it shows up in real systems.17 FebGeometric mean vs Z-scoreUnderstanding why we moved action group scoring from geometric mean to z-score normalization.15 JanFrom DS notebooks to realtime pipelines: diversity scores at scaleHow we took a data science scoring algorithm and rebuilt it for realtime distributed systems.

A few shelves for following the work: applied AI, realtime systems, product judgment, and the smaller notes I want to keep findable.

Shelves

015 articles

Applied AI learnings

Notes from turning agent workflows, personalization loops, AAmplify, FlumeWorks, and AI transformation ideas into software people can actually use.

Thompson Sampling: how explore meets exploit Geometric mean vs Z-score

022 articles

Realtime systems

Pipelines, event streams, scoring jobs, debugging trails, and the messy path from notebook idea to production behavior.

From DS notebooks to realtime pipelines: diversity scores at scale Python datetime ate all my time

032 articles

Product judgment

Product calls, scoring, experiments, uncertainty, team rituals, and the small design choices behind better decisions.

Thompson Sampling: how explore meets exploit Geometric mean vs Z-score

043 articles

Builder notes

TILs, debugging trails, onboarding notes, domain models, startup lessons, and practical things future-me should be able to find.

Audience, triggers, campaign, campaign_allocations Variants, alternates, labels, component Python datetime ate all my time

All notes

Published notes · 6 pieces

2026Thompson Sampling: how explore meets exploitA Bayesian bandit algorithm that solves the exploration-exploitation tradeoff. How it works, why it's elegant, and where it shows up in real systems.30/0310 min read

Geometric mean vs Z-scoreUnderstanding why we moved action group scoring from geometric mean to z-score normalization.17/0212 min read

From DS notebooks to realtime pipelines: diversity scores at scaleHow we took a data science scoring algorithm and rebuilt it for realtime distributed systems.15/0110 min read

2025Audience, triggers, campaign, campaign_allocationsHow audience eligibility, triggers, and campaign allocations work together in the learning pipeline.17/1112 min read

Show older notes2 more

2025Variants, alternates, labels, componentBreaking down the Composer data model: what variants, alternates, labels, and components actually mean.17/115 min read

Python datetime ate all my timeHow Python's datetime.now() combined with urllib silently breaks ISO timestamps.14/046 min read