Proposal · v3.0

Frontend Architecture

A note from the design side: V2 was an ERP demo with the intelligence bolted on. V3 inverts that — the models, their performance, and the agent are the product; the ERP is gone. Here is the proposed shape.

The argument

We cut every CRUD surface that existed only to look like an ERP. What remains is a tight loop: a fleet of thin Chronos adapters (one per product), a way to see and trust their performance, a What-If surface to interrogate them, and an agent that turns questions into verified dashboards. Six MCP tools — not twenty — so smaller self-hosted models stay reliable. The integration model mirrors Atelier: no Forge, clean LangGraph AG-UI.

Navigation · 8 pillars

Overview

Mission control — fleet health, queue, agent activity, signals at a glance.

KPI stripfleet fieldqueuedrift watch

Model RegistryHERO

The 10k-adapter fleet. Search, leaderboard, per-product detail, lineage, promote/rollback.

heat fieldleaderboardadapter detaillineage

Performance

Accuracy, backtests, drift & coverage across the fleet and over time.

backtestsdrift (PSI)coveragebaselines

What-IfHERO

Live scenario simulator over a product’s adapter — promos, price, reorders.

scenario consoleprojectionsreorder

Agent StudioHERO

Autonomous dashboard builder. Plan → MCP tools → compose → judge → ship. Currently a scripted preview; live LangGraph agent + judge planned.

tracetool logjudge rubricOpenUI board

Chat

Conversational AG-UI assistant over the same MCP toolset.

AG-UI streamtool callsinline surfaces

Signals

Meteo / fuel / inflation indices from the data-collector-proxy, correlated to demand.

signal stackcorrelationexogenous feed

Training Factory

Redis queue, Chronos LoRA jobs, EDA, MLflow registration.

queuejobsEDAMLflow

Removed from V2

Products / Suppliers / POS CRUD

Inventory · Locations · UoM

Orders · Create order · Kanban

Sales · Events · Receipts

Supply-chain flow & map

Alerts · Reports center

Forge Studio · Forge Ops

Role-heavy admin surfaces

Kept & evolved

What-If simulator — Kept — re-skinned on the instrument system, recomputes live.

Models / MLflow — Evolved into the at-scale Registry + Performance.

Exploration → Agent — Promoted to a hero, made cinematic, judge added.

Chat — Kept, moved off Forge onto clean LangGraph AG-UI.

Training — Kept as the Training Factory (queue mocked for now).

System & data flow

Source

Client BI database

read-only snapshot

Data

ETL → SupplyMind schema

daily sync · agent-shaped

data-collector-proxy

meteo · fuel · HICP

Compute

Training factory

Redis queue → Chronos LoRA

MLflow registry

10k adapters

Forecast engine

adapter inference

Serving

FastAPI BFF

REST + SSE

MCP tool layer

6 endpoints

LangGraph agent + judge

grok, self-hosted · planned

Frontend

AG-UI · OpenUI render

this app

MCP endpoints · the lean sixfewer tools → reliable small models

`get_demand_history` (product_id, days)	Daily units sold from the ETL snapshot.
`get_model_forecast` (product_id, horizon)	Calls the product's Chronos LoRA adapter.
`get_stock_position` (product_id, pos_id?)	Usable + incoming stock, shelf-life.
`get_signal_series` (signal, days)	External index from data-collector-proxy.
`get_signal_correlation` (category, signal)	Lead/lag correlation to demand.
`list_top_movers` (category, metric, n)	Ranked products by a fleet metric.

Open decisions

ETL vs query adapters

Lean ETL into a SupplyMind schema shaped for the agents, then migrate the colleague's query classes onto it. Cleaner tool contracts than adapting the raw BI DB per call.

Redis queue

Mocked in the UI for now (live progress on Overview). Swap to the real client once the library lands — interface is already isolated.

Judge model

Planned: a second grok pass with a fixed rubric + ship threshold (0.80). Semi-deterministic: components are host-approved, never executable code. The Agent Studio currently replays a scripted preview until this is wired.

Adapter rollout

MVP trains a curated subset, not all 10k. Registry already designed for the full fleet.