Learning path

Four modules that follow the transcript pipeline: understand the noisy text, engineer the LLM system, measure it honestly, then design the categories it extracts. Each lesson: concept → intuition → visual → interactive lab → realistic example → exercises → feedback → error analysis → quiz → challenge.

Not started Learning Completed Mastered

Module 1

NLP Fundamentals for Noisy Transcripts

Tokens, normalization, fuzzy & phonetic matching, dialogue structure, speakers, NER and classification — on messy Hinglish call data.

Module mastery0%

What the model actually reads — and why Hinglish transcripts can cost more to process.

25mNoisy Transcript

Text Normalization

Making noisy transcripts consistent — without destroying the signal you need.

30mNormalization

Matching “persnal lon” to “personal loan” — and knowing when a close match is the wrong match.

30mNormalization

Phonetic Matching

Rahul, Raahul, Rahool: matching by sound when spelling can't be trusted.

20mNormalization

Dialogue and Discourse Structure

A call is not a bag of sentences: turns, dialogue acts, adjacency pairs and segments.

30mNoisy Transcript

Speaker Attribution from Text

Who said it matters more than what was said: one wrong speaker label can fabricate a customer's intent.

30mSpeaker Attribution

Named Entity Recognition

Finding amounts, rates, tenures, banks and names in messy speech — and scoring it honestly.

35mNER + Extraction

Text Classification

Multi-class, multi-label, hierarchical — the label structure you choose decides what you can measure.

35mIntent Classification

Module 2

LLM Engineering

Prompts, structured output, context, grounding, judges, hallucination, sampling, model choice, caching and pipeline architecture.

Module mastery0%

Prompt Engineering

From “Is this customer interested?” to a specification the model can actually follow.

40mIntent Classification

Structured Output

JSON schemas, enums, validation and constrained generation — and why valid ≠ correct.

35mNER + Extraction

Context Engineering

What goes into the context window, in what order — and why more context can make answers worse.

35mIntent Classification

Grounding and Faithfulness

Answer + Evidence + Reasoning: every claim must point at words that were actually said.

35mEvidence Extraction

A scalable measurement instrument with known biases — validate it before you trust it.

35mLLM Verification

Fabricated entities, invented numbers, overconfident inferences — and how to catch each type.

30mLLM Verification

Sampling Parameters

Temperature, top-p and max tokens: why the same prompt can give different answers.

25mIntent Classification

Model Selection

Quality, cost, latency, context, reliability — pick the model that satisfies the constraints, then optimize.

35mIntent Classification

Prompt Caching and Batching

Stop paying full price to re-read the same 5,000-token rulebook 100,000 times.

25mIntent Classification

Multi-stage Pipeline Design

Sequential, parallel, routed: decomposition buys debuggability — and costs latency, money and compounding errors.

40mNormalization

RAG and Agents: Do You Need Them?

Most transcript-analysis problems need neither. Know when they help — and when a deterministic pipeline wins.

25mIntent Classification

Module 3

Evaluation and Data Labeling

Guidelines, agreement, metrics, imbalance, gold sets, per-stage evaluation, calibration, error analysis and significance.

Module mastery0%

Annotation Guidelines

The codebook decides your labels, your agreement, and therefore every metric you'll ever report.

Inter-Annotator Agreement

Raw agreement ≠ agreement beyond chance: Cohen's κ, Fleiss' κ and Krippendorff's α.

Classification Metrics

Precision, recall, F1, accuracy — and why the right metric is a business decision.

Macro vs Micro Averaging

80/15/5 class splits: the same predictions can look excellent or broken depending on how you average.

Class Imbalance

At 1% prevalence, a 95%-accurate detector can be wrong about most of the calls it flags.

Gold Set Design

Train, dev, test — and why tweaking prompts against your test set quietly lies to you.

Per-Stage Evaluation

End-to-end accuracy tells you THAT the system fails. Per-stage metrics tell you WHERE.

30mSpeaker Attribution

When the model says 0.9, is it right 90% of the time? Your auto-approve threshold depends on it.

30mLLM Verification

Overall F1 0.81 — but 0.62 on noisy transcripts. Slicing and root-cause analysis turn metrics into a plan.

Statistical Significance

Prompt B scored 84% vs A's 80% on 200 calls. Is B better — or did it get lucky?

Module 4

Information Extraction and Taxonomy Design

Taxonomies, open vs closed sets, embeddings and clustering for discovering and governing categories.

Module mastery0%

Taxonomy Design

Labels are a product: one dimension per field, mutually exclusive siblings, an honest 'other'.

35mIntent Classification

Open-set vs Closed-set Classification

Consistency and governance vs discovery and flexibility — and the hybrid most teams end up with.

25mIntent Classification

Embeddings and Semantic Similarity

“borrow money” and “personal loan” share no words — embeddings still put them side by side.

30mIntent Classification

Topic Modeling and Clustering

Let the data propose categories — then decide, as a human, which ones deserve to exist.

35mIntent Classification

Build a Transcript Intelligence Pipeline

Taxonomy, guidelines, labels, prompts, schema, metrics, error analysis, architecture — scored against a visible rubric.