From selling tractors to coordinating American agriculture — John Deere.

Start with one of the most familiar companies in industrial history. Deere built tractors. It ended up running the layer that American farms plan, execute, and report through. The shift from product company to coordination business is what made Deere durable — and the same shift is available to any platform with sensors at its base.

Sensor → Engine → Coordination.

Deere's path, generalized. Two transitions any analytics platform with sensors at its base eventually walks. The dashboard becomes the engine that determines outcomes. The engine becomes the coordination layer that operations run through. The rest of this deck shows what the engine looks like today and what coordination pieces are emerging.

The engine governs performance. The coordination layer regulates behavior.

These are the two distinct paths Deere walked, and they reinforce each other. The engine determines how well the system runs. The coordination layer determines how the system is run — who talks to whom, what gets reported where, when each step happens. Together they shift the locus of advantage toward the platform.

AI handles continuous reasoning. Humans intervene where it matters.

The system runs autonomously for the bulk of the analytical work — ingestion, feature derivation, drift detection, watchlist ranking, decision synthesis, audit. Proximity engineering dictates where humans are inserted: at moments closest to a consequence — confirming a root cause, approving a regime change, accepting a threshold revision. Everything else is delegated to the AI layer.

Nine lenses, three scopes — each answers a different operator question.

Every analytical output the dashboard produces falls into one of three scopes: population (what does the dataset say in aggregate), single run (what's happening in this specific batch right now), and action (what should I do about it). Lenses 1–5 build trust in the dataset. Lens 6 + 9 show what's happening in flight. Lens 7 raises the alarm. Lens 8 tells the operator what to do.

Five UX shapes over one backend — one per audience.

Same nine lenses, same endpoints, same audit trail. The deployed site exposes five distinct routes — each tuned to a different reader and a different question. Density-and-audit for analysts; route-and-ETA for operators; constellations for engineers and domain experts; particle field for the boardroom.

Trip view — Maps-style decision UI for the bioreactor floor.

Modeled on Waze. The operator gets a route, an ETA, traffic ahead, and re-route options — not a wall of metrics. Same data the analyst dashboard has, framed as navigation.

Run journey — one batch as a star, surrounded by what governs it.

For the process engineer or QA reviewer who wants the full picture of a single run. The run is a pulsing white star at the center; phases orbit blue; live readings bind to canonical parameters; regulations · vendors · failure modes fan out as outer arcs — only the slice of the parameter graph that governs this run's measurements, not the full universe.

Universe — the entire dataset as a living particle field.

For the boardroom view: every entity in the system as a glowing point. Color encodes type; particle density tracks real record counts in the dataset; connections form by canvas proximity with traveling pulses on each edge. The visual swarm is the dataset's shape.

Brain ↔ Memory — where data lives and how it flows.

The AI-first surface area sits inside the brain — per-run intelligence (regime, drift, alerts, decision) cached in a small hot tier the operator UI reads in O(1). The data lake is the historian — consulted only on cache miss, append-only for audit, and the source of population-level priors. The two feed each other on a defined cadence; the human only intervenes at proximity events (next slide).

Every table and view, with role and cadence.

Color = tier. Each card answers: what is this?, who reads / writes it?, how fresh?. The full schema lives in data/01_*.sql through data/10_*.sql and is exposed in the dashboard's methodology panel.

From observation to decision to label — the full pipeline.

Each box is a stage; the latency tag tells you the dominant cost. The hot tier ends the chain — operators read from FERM_RUN_STATE, never from raw rows.

How the brain learns from memory.

Population-level priors flow from the lake into the brain on every recompute. Per-run conclusions flow from the brain back into the lake as audit + labeled outcomes. Each pass tightens the next: the more confirmed cases FERM_OUTCOME_LABEL accumulates, the closer the differential's heuristic ★ ratings get to calibrated Bayesian probabilities.

What's autonomous, layer by layer.

Five tiers of agents, each with its own role · methods · tools · skills · memory. Sensing agents read instruments. Analytical agents fit models. Reasoning agents synthesize. Triage agents allocate attention. The narration agent translates results to plain language.

Which agent runs where, and how they hand off.

33 of 34 agents are operational today. The framework now includes 3 LLM-augmented agents (Tier 3+ / 5+) layered on top of the 30-component deterministic core, plus the original GemmaNarrator and the future-tier OutcomeLearner. One agent (OutcomeLearner) is built into the architecture as future-tier — its data pipeline is live (FERM_OUTCOME_LABEL + /api/label_outcome), the Bayesian posterior step activates once ~50 confirmed cases per cause accumulate. Everything else is in production code.

33 of 34 ACTIVE Built into 18 sensor classes · 6 analytical functions · 3 reasoning synthesizers · 2 triage agents (router + cache) · 4 LLM-augmented agents (1 narrator + 1 differential proposer + 1 confirm-by checker + 1 operator QA) · plus the planned posterior calibrator that activates on accumulated labels.

Industrial verticals the framework handles with minimum changes.

The architecture is sensor-agnostic and outcome-flexible. The nine lenses, the hot-tier inversion, the parameter knowledge graph, the agentic stack, the five presentation surfaces — none of it is fermentation-specific. Five adjacent verticals adapt with bounded engineering work; the table below shows what stays, what swaps, and the lift to ship a domain pack.