Establishing Link...
SecureEncrypted
Demo Active
Live demo — RGM Intelligence Platform v1.0
Labs · CPG revenue growth management
RGM Intelligence Platform
Four analytics lenses — descriptive, diagnostic, predictive and prescriptive — over CPG revenue data, with an AIRA analyst chat and MILP-based promo optimisation (heuristic until model training is enabled). The live demo runs entirely on synthetic data.
v1.0 · AWS EKS · live at rgm.adraca.io · demo runs on synthetic data
Interface & Capabilities
Executive summary — portfolio KPI tiles (Brand Mgr persona)
P&L bubble chart — revenue vs. margin, bubble size = volume
Model evaluation — accuracy scorecard and R-hat diagnostics
Predictive — price elasticity curve and 13-week forecast
Prescriptive — MILP optimal price table (OR-Tools SCIP)
MCTS 52-week promo calendar heatmap
Four analytics lenses
Descriptive, diagnostic, predictive and prescriptive views over the same revenue data — KPI tiles, P&L bubble charts, model evaluation, elasticity curves, forecasts and promo scenarios, filtered by persona (Brand Mgr / Analyst / VP).
AIRA analyst chat
A LangGraph-orchestrated analyst on AWS Bedrock Nova Pro, streamed over SSE. It answers with real tool calls into the platform — elasticity lookups, MILP recommendations and forecasts — with Presidio PII redaction in the middleware.
MILP + MCTS promo optimisation
OR-Tools SCIP solves optimal prices against floors, ceilings and elasticities in ~13 ms; an MCTS (UCB1) search builds a 52-week promo calendar; scenarios go through a DynamoDB approve/reject workflow. Elasticity and forecasts are heuristic until model training is enabled.
Architecture
Bayesian Demand Model
SageMaker-backed training path with R-hat convergence diagnostics read from MLflow; runs on a heuristic fallback until model training is enabled.
Price-Elasticity Estimation
Per-SKU elasticity matrix served from the Athena gold layer; demo values stand in until training populates the store.
MILP Optimiser (OR-Tools SCIP)
Optimal price solve under floor/ceiling constraints, ~13 ms measured; results feed a DynamoDB approval workflow.
MCTS Promo Calendar
UCB1 tree search over promotional slots producing a 52-week calendar, chainable with the MILP solve in one endpoint.
Stage 01
Ingest
REST JSON and EDI X12 852/867 payloads accepted into the S3 bronze layer; nightly Airflow DAG picks them up.
Stage 02
Harmonize
Bronze → Silver → Gold medallion pipeline via Glue ETL; SKU entity resolution with Levenshtein matching and Qdrant embedding search; gold layer queried through Athena.
Stage 03
Model
Bayesian/DML demand modelling with R-hat diagnostics via MLflow — elasticity and volume forecasts fall back to heuristics until training is enabled.
Stage 04
Optimize
MILP optimal prices and the MCTS 52-week promo calendar; AIRA surfaces the results conversationally.
Deployment Specifications
Analytics Lenses
Descriptive · Diagnostic · Predictive · Prescriptive
Optimiser
OR-Tools SCIP MILP (~13 ms measured) · MCTS (UCB1, 52-week)
AI Analyst
AIRA — LangGraph on Bedrock Nova Pro, SSE streaming
Data Layer
S3 medallion lakehouse · Glue ETL · Athena · Qdrant SKU vectors
Hosting
AWS EKS
Demo Data
Synthetic (seeded generator) — flagged with a DEMO DATA badge
adraca --log-stream
[--:--:--]INGEST ACCEPTED [EDI X12 852 -> S3 BRONZE]...
[--:--:--]MEDALLION PIPELINE [BRONZE -> SILVER -> GOLD]...
[--:--:--]ENTITY RESOLUTION [LEVENSHTEIN + QDRANT VECTORS]...
[--:--:--]ELASTICITY LOOKUP [ATHENA · HEURISTIC FALLBACK]...
[--:--:--]MILP SOLVE [OR-TOOLS SCIP · ~13MS]...
[--:--:--]MCTS CALENDAR [UCB1 · 52 WEEKS]...
[--:--:--]AIRA RESPONSE STREAMING [NOVA PRO · SSE].