Confidential · Draft · Jun 2026
System Architecture · Flow

LumicIQ — end-to-end flow

One canonical event stream → a learned attention gate → an LLM agent reasoning over composed context & scoped knowledge → a routed, quantified recommendation to the right person. Every call is audited; outcomes feed back to improve the prompts.

Source Spine / projection Detect Agent / LLM Deliver Storage data / control flow learning feedback
events proposed archive prompt layers LLM call RAG state + memory LlmCallRecord weekly · prompt evolution 01 · SOURCES Sensors & adapters POS · traffic · inventory weather · video · simulator 02 · SPINE Event Hubs RelayEvent · ULID idempotent partition hash(tenant, store) 03 · ATTENTION Interestingness learned gate · no per-tenant config 03 · FLOOR SafetyNet ~10 hardcoded critical conditions 04 · REASONING The Agent StoreAgent.single@1.0 · per (tenant,store) composes the 5 context layers + NOW validates → suggestion or answer writes memory + audit CONTEXT Composed context VerticalPack · Tenant · Store LLM ILlmProvider OpenRouter · per-task models 05 · ROUTE Router recipient + channel · escalation 06 · CHANNELS Deliverers SignalR · push · earpiece · sms 07 · HUMAN Client · React Native act · dismiss · delegate · ask PROJECT StateProjection emit-then-project · idempotent OPERATIONAL Stores Cosmos · Redis · Azure SQL current truth · memory · config KNOWLEDGE Vector index · RAG Azure AI Search platform → enterprise → store AUDIT Immutable record Blob + ADX every prompt + response · SOC 2 08 · LEARN Learning loop outcomes → backtest → prompt evolution → ship via experiments