Pranav's learning log

Notes from building, breaking, and learning.

AI systems, developer tools, and the experiments between them.

Draft

Build note

Building Tweakler: cloning interfaces without guessing

What changes when a design reference becomes structured evidence instead of a screenshot prompt.

  • Agents
  • Design systems
  • Tooling
Coming soon
Draft

Field note

The missing quality layer in coding-agent evaluations

Notes on separating policy performance, captured evidence, and the judgment that connects them.

  • Evaluation
  • Coding agents
  • Benchmarks
Coming soon
Draft

Learning log

Training GPT-2 on 10 billion tokens from scratch

The implementation details that made the jump from a readable transformer to a trainable one.

  • Foundations
  • Transformers
  • Training
Coming soon
Draft

Architecture

Routing sensitive data before it reaches the cloud

A practical model for classifying PII and choosing between on-device and cloud inference.

  • AI systems
  • Privacy
  • Inference
Coming soon
Draft

Product note

Designing an AI copilot for sensitive case notes

What early-childhood assessors taught me about structured outputs, trust, and useful automation.

  • Design
  • Healthtech
  • Human factors
Coming soon
Draft

Research note

When product analytics can adapt the interface

Exploring the boundary between observing user behavior and automatically changing a product.

  • AI systems
  • Analytics
  • Adaptation
Coming soon