Andrew VernonResume
← All work

In development · 2026

EDGAR Events

SEC 8-K filings turned into typed, entity-resolved event rows, with evals and a cost ledger you can audit.

FORM 8-K / ITEM 1.01event_typecounterpartyamounteffective_date
Role
Design, build, evaluation
Year
2026
Stack
Python, Claude API, AWS Step Functions, Lambda, DuckDB, Terraform
Links
Repository not public yet
$15
Hard monthly cap, cloud and model APIs combined
~2,700
Filing items per month, measured
4
8-K item types in scope
10/s
SEC fair-access request ceiling

Data flow

  1. SEC EDGAR8-K header and primary document
  2. Bronzethrottled fetch, replayable
  3. Contract gateschema checks, quarantine path
  4. LLM extractionkeyed by content, model and prompt version
  5. Entity resolutionrules, embeddings, then LLM
  6. Evals and datasetgolden set, cost and freshness lines

Problem

Companies announce material events in 8-K filings within days: new agreements, officer changes, bankruptcies. The facts sit in narrative HTML, with no structured tags to read them from. Vendors charge thousands of dollars per seat for the structured version.

The goal is a public dataset, refreshed daily, where each 8-K item becomes typed rows (event type, dates, counterparties, amounts, named officers), filers and counterparties resolve to canonical entities, and every published number carries its accuracy and its cost. All of it has to run under $15 a month.

Approach

Versioned, idempotent enrichment. Each extracted row is keyed by a hash of the content, model, prompt version, schema version and parameters. Bumping a prompt marks only the affected rows stale, and the backfill is selective.

Entity resolution in tiers. Rules and probabilistic matching handle the easy pairs. Embedding blocking narrows what is left. An LLM adjudicates only the uncertain band, and the result is reported as F1 and dollars per 1,000 pairs.

Evals before trust. A hand-labeled golden set scores value accuracy with confidence intervals. A judge regression test and a forced model-swap diff catch drift when a model or prompt changes.

A cost ledger at list price. Every model call lands on the ledger at list price, even when a credit pays for it, so the published cost line reflects what the pipeline would cost anyone.

Contracts and lineage. A contract gate with a quarantine path sits between bronze and extraction. Lineage is emitted as OpenLineage events with a custom facet for LLM calls.

Serverless on a schedule. EventBridge triggers Step Functions, which runs Lambda steps over Parquet in S3. DuckDB builds the rollups and CloudFront serves a static dashboard.

Result

The SEC client, throttle, bronze store, item parser, enrichment keys and cost ledger are built and tested. A live run against EDGAR measured about 2,700 in-scope items a month across items 1.01, 1.03, 5.02 and 8.01, which sized the model budget. Item 2.02 was dropped after measurement showed its text is a stub pointing to an exhibit.

Next: reference extractions with a strong model, the golden set, judge calibration, then production hardening on AWS.

Next projectOnchain Indexer