LLMs extract people, companies, models, audience level, and content depth — metadata that would be tedious to create manually.

Semantic Enrichment

What Semantic Enrichment Means

The decorated files contain metadata that wasn’t in the original emails. This metadata — who’s mentioned, which companies appear, what models are discussed, who the audience is, how deep the content goes — is extracted by the LLM from the content itself.

This is the Semantic Enrichment pattern: using LLMs to add structured meaning to unstructured content. It’s one of the most valuable things AI agents can do.

What Gets Extracted

People (with Roles and Affiliations)

people_mentioned:
  - name: "Andrej Karpathy"
    affiliation: "OpenAI (former)"
  - name: "Andrew Ng"
    affiliation: "Landing AI"
  - name: "Fei-Fei Li"
    affiliation: "Stanford"
  - name: "Holden Karau"
    # No affiliation listed — agent can leave it empty

The agent identifies people from context: names in attributions, quoted speakers, people referenced in article text. It often captures roles and affiliations too, even though those aren’t always explicitly stated in the same sentence.

Companies

companies_mentioned:
  [
    "Cursor",
    "Hugging Face",
    "OpenAI",
    "Anthropic",
    "Meta",
    "Google",
    "Microsoft",
    "Aleph Alpha",
  ]

Notice this includes companies of all sizes — from major labs (OpenAI, Google) to startups (Cursor, Aleph Alpha). The agent doesn’t need a predefined list; it recognizes company names in context.

AI Models

models_mentioned:
  ["GPT-4o", "Pharia-1-LLM", "Phi-3.5", "Hermes 3", "Jamba-1.5", "Gemini"]

Models are often mentioned in passing — “OpenAI’s GPT-4o,” “Meta released Llama 3,” “a new paper benchmarks DeepSeek-V3.” The agent extracts them from the surrounding text, handling variations in naming conventions.

Content Characteristics

audience: mixed
depth: overview
has_research_papers: true
has_code_examples: false
is_premium: false

These are subjective judgments that the LLM is surprisingly good at:

Audience: Is this written for ML researchers (technical), business leaders (business), general readers (general), or a mix (mixed)?
Depth: Is this a high-level summary (overview), a moderately detailed piece (intermediate), or a comprehensive deep-dive (deep-dive)?
Research papers: Does the content reference specific academic papers?
Code examples: Are there code snippets in the content?
Premium: Is this premium subscriber content?

How the Agent Does It

The classification agent reads the entire markdown file (often 2-8KB of text) and:

Identifies the series from the subject line (regex + emoji detection)
Determines content type from structure (multiple sections = digest, Q&A format = interview, etc.)
Assigns a primary topic based on overall theme
Extracts entities (people, companies, models) by scanning for known patterns and context
Judges audience and depth from writing style and technical level
Generates tags from key concepts discussed
Sets boolean flags (has papers, has code, is premium) from content inspection

All of this happens in a single pass. The agent doesn’t need multiple API calls or a pipeline of specialized extractors — one LLM call per file handles everything.

The Power of This Pattern

Scale

Doing this manually for 210 files at even 5 minutes each would take 17.5 hours. The agent does it in ~15 minutes (a 70x speedup).

Consistency

Humans get tired, inconsistent, or biased after classifying dozens of items. The agent applies the same taxonomy rules to every file. The 3 validation failures were due to malformed source YAML, not classification errors.

Richness

A human classifier might tag a few topics and call it done. The agent reliably extracts:

430 unique people with affiliations
553 unique companies
615 unique model mentions
Hundreds of content-specific tags

This creates a cross-reference graph — from any article, you can navigate to every person, company, and model it mentions.

Discoverability

Before classification, finding content meant scanning filenames or full-text search. After classification, you can query:

“Show me all deep-dives about agents from 2024”
“Which articles mention Andrej Karpathy?”
“What are the top companies discussed in FOD digests?”
“List all technical explainers mentioning Llama models”
“Show interview transcripts with audience=mixed”

The frontmatter becomes a queryable database.

The Cross-Reference Network

The extracted metadata enables rich cross-linking:

Article: "Golden Age for Indie Devs"
    mentions → Andrej Karpathy, Andrew Ng
    mentions → Cursor, Hugging Face, OpenAI
    mentions → GPT-4o, Phi-3.5, Hermes 3
    ↓
Entity page: "Andrej Karpathy"
    appears in → 34 articles
    ← links to → OpenAI, Tesla
    ← connected to → Andrew Ng, Fei-Fei Li
    ↓
Entity page: "OpenAI"
    appears in → 133 articles
    ← connected to → Anthropic, Google, Microsoft

This is possible because the agent consistently extracts and normalizes entity names across all files.

Why LLMs Are Uniquely Good at This

Traditional NLP approaches to entity extraction (spaCy, NER models) can identify people and organizations but can’t:

Disambiguate context: Is “Anthropic” a company or an adjective?
Recognize novel entities: A new startup mentioned for the first time
Identify AI model names: “GPT-4o” or “Pharia-1-LLM” aren’t in standard NER training data
Judge content characteristics: Audience level and depth require reading comprehension, not pattern matching
Handle varied formats: Podcast transcripts, digests, explainers, and profiles all have different structures

LLMs handle all of these naturally. They understand that “Cursor” in an AI newsletter is a company (not a UI element) and that “Hermes 3” is a model release (not a Greek god).

Next: Phase 4: Validation — Scripts, edge cases, and iterative refinement.