LumicIQ — what we're building, end to end

On-premise connectors, running on a store computer, that turn systems the platform doesn't own — point-of-sale, labor, inventory, BI — into sensor streams without vendor integration work. They attach to the operator's already-authenticated browser session (no stored credentials), navigate deterministically, and use an LLM for exactly one step: extracting a declared set of metrics into a typed schema. Readings that fail schema validation are never emitted.

Two capture modes ship today: a dashboard connector that walks the URL-addressable report views of a web back-office (site adapters make each new vendor a small config; Loyverse is the first), and a screen connector — a human-in-the-loop companion where the operator shares a tab and the connector reads the numbers on it. Both emit structured, idempotent events onto the same spine as every physical sensor; all reasoning stays with the platform Agent.

A per-store perception unit (ADR-0009) that runs the vision at the edge — a store box, a customer datacenter, or a cloud-hosted instance — and sends only derived data across the trust boundary. Deterministic code does the perception: object detection (ONNX / YOLO11), multi-object tracking (ByteTrack), and appearance re-identification (CLIP galleries) that carries a person's identity across track loss and camera hand-off. From the tracked world it emits typed video events (a shopper lingered, an aisle is crowded, someone is waiting unassisted) plus a periodic world-state aggregate — occupancy and dwell per zone — onto the spine.

Crucially, no VLM runs at the edge. Expensive visual "meaning" is a cloud cascade: the edge ships trigger evidence and rolling 60-second video segments to a cloud world-store, and the platform re-inspects the actual pixels only when a question or a shelf-level judgment makes it worth the spend. Perception at the edge, meaning in the cloud — the Overwatch deck is the full-pipeline reference implementation this port is derived from.

Whatever the store already operates — IoT sensors, people counters, inventory feeds — plus external context such as weather and local events, pulled by per-source adapters with deduplication, rate limiting, and health monitoring. The platform is deliberately sensor-agnostic: it standardizes signals, it does not prescribe hardware.

A digital-twin store built in Unreal Engine 5 — inventory, shoppers, and scripted scenarios with known ground truth. Its virtual camera rig and event generators speak the exact same contracts as production sensors, so the entire loop — video → events → attention gate → agent → human — can be trained, exercised, and regression-tested synthetically. Moving to physical cameras is a configuration swap, not a rebuild.

All signals become one canonical event stream — idempotent, deduplicated, partitioned by (tenant, store), with every consumer seeing every event. Everything downstream is replayable from this stream, which is what makes the audit and learning loop possible.

A universal, learned statistical attention gate — per store / event type / time-of-week baselines, with no per-customer rules — that decides which events merit expensive LLM reasoning. This is the cost-control mechanism that makes continuous LLM supervision of a high-volume physical event stream economically viable.

Beneath it sits the SafetyNet: a small deterministic floor of critical conditions the platform must never miss, independent of both the learned gate and the LLM.

A persistent LLM reasoner per (tenant, store). For each gated event it composes a five-layer context — platform → vertical → tenant → store → situation — plus the live "now", reasons over current state, its own memory, and scoped knowledge, and emits either a quantified suggestion or a conversational answer.

There is no rules engine anywhere: customer configuration describes the business, it never encodes rules. Behavior lives in composed context, so it can evolve without code changes.

Every LLM call — the exact composed prompt, layer versions, retrieved context, and response — is stored as an immutable, reconstructable record. Every suggestion's disposition (acted on, assigned, dismissed, ignored, expired) joins it, scoring how useful each alert actually was, so candidate prompt changes are backtested against recorded history and shipped through experiments. The product improves weekly as measured context edits, not deployments.

The router resolves who should hear about a suggestion and on which channel, with escalation policies — and never edits agent-authored content. Specified channels include in-app, push, and SMS; two audio endpoints are planned — a wearable earpiece and a walkie-talkie bridge (text-to-speech onto the radios many stores already carry) — both treated as priority-preemptive serial audio media: critical interrupts lower-priority speech, queues are depth-limited, and timed fallback chains hop channels (audio → push → SMS) until someone is reached.

A mobile-first floor app: a live queue of suggestions with quantified impact and the agent's reasoning on tap — act, assign to an associate, dismiss, or hand back — plus a direct conversation with the store's agent.

Every suggestion's disposition is recorded — acted on, assigned, dismissed, ignored, expired — alongside the reasoning record that produced it, so the usefulness of every alert is measurable per store, per event type, and per prompt version, and feeds directly into adjustment and learning.