Touchline

A marketing analyst in a box: business question in, analyst-grade SQL, deterministic attribution, and a cited interpretation out - or an honest refusal when the evidence does not support a claim.

Every number about credit and lift is computed by deterministic Python against a real warehouse. The LLM investigates, explains, and cites; it never invents a figure. Same discipline as BondLens, applied to the domain Epsilon's Account Analytics team lives in - where the team already runs AI agents that generate analyst-grade SQL and treats evaluation harnesses as part of the job.

Status

Wave 0 of 6 - the deterministic core:

• Attribution models (last-touch, first-touch, linear, time-decay) with a conservation contract: credit assigned sums to revenue converted, asserted on every run
• DuckDB warehouse schema (spend, conversions, ordered touchpoints) with atomic inserts and path streaming
• Seeded synthetic-data generator with planted channel multipliers, so interpretations can be graded against a known ground truth

Planned next: text-to-SQL analyst CLI with cost ledger (wave 1), evaluation harness + geo-lift incrementality (wave 2), BigQuery sandbox + Databricks Free Edition adapters (wave 3), one recorded cloud-LLM run with a local-vs-cloud comparison table (wave 4), agent tool calls governed by Turnstile (wave 5).

Run the gates

uv sync
uv run pytest
uv run ruff check src tests
uv run mypy src        # strict

Generate the demo warehouse

uv run python scripts/generate_synthetic.py data/marketing.duckdb

Seeded and deterministic: the same command produces the same warehouse, and the generator's README section documents the planted effect sizes.

Honest scope

A portfolio project, local-first, not production software-as-a-service. The synthetic warehouse is a stand-in for real campaign data; the identification assumption behind any lift estimate is stated next to the estimate, and a question the data cannot answer gets a refusal, not a confident guess.