VOL.01 HUB VOL.02 PORTFOLIO
LIVE · BENGALURU EST. 2024
★ ALL EXPERIMENTS · ARCHIVE

THE ARCHIVE

Every experiment, in reverse chronological order.

1RgRAG3
2EmEmbed0
3RkRerank0
4EvEvals3
5ClCalibrate0
6BmBenchmark0
7MlMLOps2
8LtLatency0
9CsCost0
10ScScale0
11PrPrompts0
12SfSafety0
13DbDebug0
14NtNotes1
15ThTheory1
16OpOptimise2
★ 12 experiments
⚠
EXP #012⚠ Learning
How Do You Grade a RAG System?
Retrieval found the right passage — at rank 80, where the model never reads it. Classification metrics can't grade a ranked list. Building MRR and NDCG from scratch, and the one idea that makes NDCG click: rank is the guess, grade is the truth.
2026-06-2910 minrag-pipeline-build
A relevant chunk showing up in the retrieved set means retrieval did its job — presence is what matters.
⚠
EXP #011⚠ Learning
Pick the Wrong Metric and You Go Blind
A model has no single "performance" — it has a confusion matrix, a score, and a dial. Every metric is a slice of those four cells, and the wrong slice hides the exact failure you shipped. A bouncer, a club, and the 0.99 that lies.
2026-06-2912 minml-foundations
Accuracy (or AUC) tells you how good a classifier is — read the number and you know where you stand.
⚠
EXP #010⚠ Learning
Cross-Validation: Why a Great CV Score Lies
A 0.92 in cross-validation, a 0.71 in production, and no code changed. Leakage and the winner's curse inflate CV without touching the real world — and the fix is treating CV as a strict simulation of deployment.
2026-06-1410 minml-foundations
A high cross-validation score means the model will generalise — trust the number and ship.
⚠
EXP #009⚠ Learning
Regularisation: Four Knobs That Aren't the Same Knob
I assumed L1, L2, dropout and BatchNorm were one family of regularisation knobs — pick whichever. Far from it: BatchNorm is an optimiser, L1 and L2 shrink by different machines, and "L2 inside Adam" isn't even weight decay.
2026-06-149 minml-foundations
L1, L2, dropout and BatchNorm are interchangeable regularisation knobs — same goal (shrink the train/val gap), so reach for whichever is handy.
✓
EXP #008✓ Achievement
Monitoring in Production: When the Dashboard Is Green and the Model Is Dying
Your inputs are stable, your predictions are stable, every monitor is green — and the model is quietly getting worse. How that happens, and the four things you watch so you see it before your users do.
2026-06-127 minproduction-ml
A production monitoring system must combine label-free signals (available day-one) with delayed-label performance tracking via cohorts — because distribution monitors are structura…
✓
EXP #007✓ Achievement
Detecting Drift: PSI, KS, and the Stat That Lies When You Bin It Wrong
Three kinds of drift, four ways to measure it, and the one detail — your binning — that decides whether the alarm fires on noise or stays silent through a real shift.
2026-06-126 minproduction-ml
Distribution shift between training and production is measurable without labels, and the right statistic depends on the data type — but every binned metric (PSI, chi-square) can be…
✓
EXP #006✓ Achievement
Optimisers: From SGD to Adam, One Fix at a Time
Every optimiser is a patch on the one before it. Learn the chain — momentum fixes zigzag, RMSprop fixes the dying learning rate, Adam fuses both — and you can derive any of them from memory.
2026-06-079 minml-foundations
The optimiser family is not a list to memorise but a chain of fixes: each method addresses a specific failure of its predecessor, and Adam is exactly Momentum (direction) plus RMSp…
✓
EXP #005✓ Achievement
Bias-Variance: The U-Curve Everyone Cites and Few Can Diagnose
Anyone can label a model "overfit". The signal is reading it off two numbers, knowing which lever to pull first — and knowing exactly where the classical theory stops being true.
2026-06-078 minml-foundations
Test error decomposes cleanly into bias, variance, and irreducible noise — and a model's failure mode can be diagnosed from the gap between training and validation error, then trea…
✓
EXP #004✓ Achievement
CRAG: Grade Your Retrieval Before You Trust It
Retrieval can fail silently. CRAG adds a confidence-scoring evaluator between the retriever and the generator — discarding bad chunks before the LLM ever sees them.
2026-05-277 minrag-pipeline-build
Scoring retrieved chunks as CORRECT / AMBIGUOUS / INCORRECT before generation, and routing to a fallback when quality is low, will reduce hallucination caused by bad context more r…
✓
EXP #003✓ Achievement
Step-Back Prompting: Retrieve the Principle Before the Detail
Before answering a specific question, ask a broader one — retrieve foundational context, then answer the original with both in hand.
2026-05-276 minrag-pipeline-build
Abstracting a specific query to its underlying principle and retrieving on that abstraction will surface foundational context that improves answer quality on complex, multi-step qu…
✓
EXP #002✓ Achievement
HyDE: Embed the Answer You Wish You Had
When your query and its answer live in different parts of the embedding space — and how generating a hypothetical answer first bridges that gap.
2026-05-276 minrag-pipeline-build
Embedding a hypothetical answer to a query will produce a vector closer to relevant documents than embedding the raw query directly, because answers and answers share vocabulary in…
✓
EXP #001✓ Achievement
Vector Spaces & Cosine Similarity: The Math Behind Dense Retrieval
Why two texts can mean the same thing without sharing a single word — and how a dot product turns that intuition into a retrieval engine.
2026-05-235 minrag-pipeline-build
Semantic similarity between texts can be captured as geometric proximity in high-dimensional space, making cosine similarity a principled and interpretable retrieval signal.