Agent 08 · Retail
DemandCast Agent
Watches forecast accuracy and replaces failing forecasts with ML forecasts.
Narrow scope firstInsightLangGraph
Found3
Fixed + verified3
Held for approval0
Recalln/a
One offline run on demo data with planted problems (recorded platform test, November 2026).
How DemandCast heals data, step by step
INData inSame inventory dataset; holiday calendar.
01Detect3 problems found
02Diagnose1 problem type named
03Plan3 safe to fix · 0 need a person
04Fix + verify3 fixed and re-checked
05ReportPlain-English report and Control Tower record
- Client problem
- Forecasts silently go stale; buyers stop trusting them.
- What it does
- Measures the existing forecast's error per product over the last 14 days and flags drift above 25% MAPE.
- How it heals
- Publishes a 7-day ML forecast (gradient boosting on lags, price, promotion) for every store of each drifting product.
- Human gate
- None for the forecast; order changes go through InventoryIQ.
- Typical buyer
- Retail chains (planning team)
- Built with
- LangGraph self-healing loop
- Demo data
- Same inventory dataset; holiday calendar.
- Tables
- inventory ← Kaggle anirudhchauhan/retail-store-inventory-forecasting-dataset
- Target outcome
- Target: forecast error (MAPE) back under the agreed limit after each retrain.
- Path to production
- Reposition as forecast health monitoring until a client provides 12+ months of history.
Issues detected on the demo data
| Issue | Risk | Proposed fix | Count |
|---|---|---|---|
| forecast_drift Legacy forecast MAPE 43% over last 14 days (limit 25%) | medium | Replace with ML forecast for the next 7 days | 3 |
How we demo it
- Show the legacy forecast error per product.
- Run DemandCast: drifting products are flagged.
- Open gold/demand_forecast with the new ML forecasts.
Built on AWS
Runs in the platform container on EC2 Graviton (or on demand in Lambda), writes healed tables to the S3 lakehouse, uses Amazon Bedrock only to write explanations, and reports six metrics to CloudWatch.
Step into DigiQuench.
Agents that heal your data.
Watch an agent find planted problems, fix the safe ones and wait for your approval on the rest. Thirty minutes, live.