An AI co-pilot for boring ops work in retail teams.
Ingest orders & customer data, score churn and demand, recommend the next action, and dispatch to Slack, email, or Sheets — every plan graded against a 55-case eval corpus.
› which 5 products bring in the most revenue? plan: top_products(n=5) trace: top_products → 5 rows in 0.04s answer: · SKU-247 $182,431 · SKU-019 $151,902 · SKU-118 $138,544 · SKU-882 $129,107 · SKU-014 $118,392 › which customers are about to churn? plan: churn_risk(days=60) trace: 1,204 customers · 18% At-Risk · 7% Churned
Detect, recommend, and ship — eval-graded.
Two product surfaces share a single tool registry. Both code paths are scored on the same 55-case corpus, so improvements show up everywhere a tool is used.
Detect → Recommend → Act
Watches your KPIs, flags incidents with anomaly + change-point detection, and ships the right next step to Slack, email, or Sheets. A Thompson bandit learns from your feedback.
- · Anomaly + drift detection per metric
- · Bandit-driven recommender across 5 action templates
- · Slack / email / Google-Sheets dispatch
- · Bcrypt + Fernet encrypted credentials
Ten tools, eval-graded
Plain-English questions over orders & customers. Plans are scored on tool-match, args-match, forbidden-tool avoidance, and must-succeed execution.
- · Revenue trends · top products · top customers
- · RFM segments · BCG quadrants · co-purchases
- · Price elasticity · churn risk · cohort retention
- · Heuristic + Gemini backends, both eval-tested
Two corpora, three numbers.
The headline 100% is the heuristic against questions it was tuned on — useful for catching regressions, not for proving generalization.
The held-out split paraphrases each question with vocabulary the heuristic has never seen, so we can see how each backend behaves on novel input.
Both numbers are real and pinned in tests/fixtures/.
Regression-catcher. Same regex, same questions — moves only when something breaks.
"Identify my whales", "bucket my buyers" — none of the regex's keywords appear. The heuristic falls off a cliff.
Same paraphrases through Gemini. ~4× the heuristic — meaningful generalization, with room to grow.
Nine stages. All observable.
Every stage is restartable, instrumented, and eval-checked.