Lattice · Cross-System Intelligence · Support & Operations
The operating layer that sees across every system
One intelligence layer over 14 production databases, Zendesk, Jira, Confluence, Planhat and PostHog — two reasoning cores that reconcile before anything reaches a decision, two specialist fleets, a writable operating board, four live dashboards, and 16 automations. Status is honest: live is verified live; provisioned and degraded are labeled, never dressed up.
Live — verified running
Provisioned — dispatch-ready
Degraded — runs, delivery reconnecting
live signal · 8 systems → 1 intelligence layer → 2 reconciling cores
The reasoning core
Signal core Live
Reads the raw source across every system
Pulls from every system and reconciles what's actually true — then builds the surfaces, runs the fleet, and produces the numbers. Answers "is this what's really happening?"
Strategy core Live
Reads independently, then reconciles
Forms an independent read of what to do, then reconciles against the signal core. Answers "is this the right call?" Divergence is surfaced, not silently resolved.
The fleet — live operating rack
Support fleet
6 specialist agents · ticket-review pipeline running Canopy intake every 10m · TechOps · QA · +3 dispatch-ready
1 live · 5 ready
Operations fleet
Knowledge base + codebase Q&A running every 2m across 398 repos · weekly review · monthly Confluence refresh
2 live · 1 ready
The Loop
Writable operating board · Airwallex + projects · the cores and fleet hand off here — no human courier
live
Dashboards
4 surfaces · Support Intelligence Center (7-tab) · Support Ops · Program Mgmt · Metrics
4 live
Automations
16 crons · Plaid pulse → iMessage · Canopy intake · backlog pulse · dispatcher · nightly self-maintenance
13 live · 3 reconnecting
Transcript intake
Read.ai meeting transcripts → extract action items → post to the right board. Anything it can't place with confidence is flagged for human review, never guessed.
in build
The shared stack
Runtime Hermes Agent · macOS local · DigitalOcean · Cloudflare · cron + dispatcher
Reasoning Claude · MCP servers · prompt caching · persistent memory + skills
Systems of record Metabase (14 DBs) · Zendesk · Jira · Confluence · Planhat · PostHog
Intake & collab The Loop / Kanban · Slack · iMessage · Read.ai · Google Workspace
Code Git / GitHub (398 repos) · codebase Q&A · skill library · Python · SQL
Security per-profile vault · VPN-gated data · FileVault · Cloudflare Access
The guardrails — why the output can be trusted
Sourced and labeled
Every finding carries a confidence tier and a named source, and fact is labeled separately from interpretation — you can challenge the read without doubting the data.
G1 verified primary · G2 multi-source · G3 single source
G4 inference · G5 extrapolation
Two independent reads
The signal core reads what's true; the strategy core forms an independent read of what to do; the two reconcile before anything reaches a decision. When they diverge, the disagreement is surfaced — never silently resolved.
Read / write boundaries
Systems of record are read across freely; writes are deliberate and scoped. Sensitive data stays VPN-gated and per-profile, and privacy boundaries between agents are enforced by design, not convention.