GET
EXCITED.
∇ thinking =
build
·
break
·
document
·
repeat
Every post is an experiment. Hypothesis first, observations next, verdict at the end. Achievements and learnings are equally first-class — because showing the wrong turns is what builds trust in the map.
★ CURRENTLY IN THE LAB
Active Experiments
Total experiments
17
Achievements
8
Learnings
9
Active quests
5
1RgRAG3
2EmEmbed0
3RkRerank0
4EvEvals3
5ClCalibrate0
6BmBenchmark0
7MlMLOps2
8LtLatency0
9CsCost0
10ScScale0
11PrPrompts0
12SfSafety0
13DbDebug0
14NtNotes1
15ThTheory1
16OpOptimise2
17AtAttention4
18FtFine-Tuning1
EXP #017
Fine-Tuning: When Not To, and How
The instinct is: to make an LLM do something new, retrain it. Almost always wrong. Fine-tuning teaches behaviour, not facts; the cheaper rungs (prompt, RAG) solve most problems; and when you do fine-tune, it is a low-rank LoRA adapter behind eval gates, not a full retrain. The LoRA math, the SFT→RLHF→DPO ladder, the production traps that sink naive fine-tunes, and a decision framework — with the interview scenarios where people get it wrong.
★ ACTIVE QUESTS
Q-01
Building RAG from Scratch
Personal knowledge retrieval — from zero to state-of-the-art
✓ 0 achievements ⚠ 0 learnings 5 experiments
2026-05 → present →
Q-02
ML Foundations from First Principles
The fundamentals every MAANG interview tests — owned, not memorised
✓ 0 achievements ⚠ 0 learnings 5 experiments
2026-06 → present →
Q-03
Production ML — Monitoring, Serving, Scale
Keeping models alive after launch — the MLOps half of the job
✓ 0 achievements ⚠ 0 learnings 2 experiments
2026-06 → present →
All quests →
— every post is an experiment. test the test before you test the model. —