Mando Platform

Meet Mando — your AI operations team.

A team of 13 AI specialists working around the clock for Increase Construction. They handle your ads, leads, content, meetings, admin, and systems — so you and Drew can focus on clients and strategy.

Agents on
the team

Skills &
capabilities

Systems
running

24/7

Always
on

Secrets
secured

What does Mando actually do for the business?

Mando isn't one AI — it's a coordinated team of specialists, each with a defined role and access to the right tools for that role. Think of it as having a fully staffed ops and marketing department that never sleeps, never forgets, and costs a fraction of the alternative.

📣

Keeps the content pipeline moving

Every Facebook ad, social post, and SEO article gets reviewed by Elara before it goes live. Nothing hits the public that isn't on-brand and accurate. The pipeline runs daily — no manual trigger needed.

🎯

Fills the BD pipeline automatically

Boba finds target businesses, researches them across multiple sources, and creates enriched leads in Pipedrive — complete with outreach drafts ready to send. 50 new leads a day, hands-off.

👷

Routes every ECTC job application

Every inbound application to ECTC is automatically classified by location and routed to the right regional coordinator. No manual triage. No misrouting. Greg's team just receives the right leads in the right inbox.

📝

Captures every meeting decision

Scribe joins Zoom calls, records them, and produces a structured summary — decisions made, commitments given, action items with owners — within minutes of the call ending. No more "what did we agree on?"

⚙️

Watches the systems so you don't have to

Rex monitors all 12 services every 2 hours. If an API token expires, a service crashes, or costs spike — Rex flags it in Slack before it becomes a client-facing problem. You don't have to check dashboards.

🏆

The team gets better every week

Ahsoka reviews every agent's work weekly, writes specific feedback into their memory, and holds them accountable to it. Agents that keep making the same mistake get escalated to Sean. The team improves systematically.

How does Mando fit into the IC tech stack?

Every piece of software IC uses is wired into Mando. Agents can pull live ad performance from Meta and Google, create deals in Pipedrive, post to Facebook/LinkedIn/Instagram, send emails through Gmail, read your Google Calendar, and produce reports — all without Sean or Drew having to touch a keyboard.

Platform	What Mando does with it	Which agents
Meta / Facebook Ads	Reads live campaign performance, frequency, creative analytics. Elara uses this to review and adjust ad strategy.	Elara, Cid
Google Ads	Reads campaigns, keywords, quality scores, search terms. Feeds into reporting and strategy.	Cid, Elara
Pipedrive CRM	Creates and enriches leads, updates deal stages, logs outreach notes automatically.	Boba
Facebook / Instagram / LinkedIn	Generates, reviews, and publishes social posts on a daily/weekly schedule.	Elara, Wren
Google Calendar	Reads upcoming meetings, sends reminders, preps briefings before calls.	Cara, Fennec
Gmail	Drafts and sends emails on behalf of Drew and Sean. Meeting follow-ups, client comms, digests.	Cara, Fennec, Scribe
Zoom	Joins meetings, records audio, produces transcripts and structured summaries.	Scribe
WordPress (IC website)	Publishes SEO articles, updates content, manages Google Business Profile posts.	Luca (automated)
Serper + Apollo + ABN	Multi-source lead enrichment — company data, contacts, ABN verification, website scrape.	Boba

The Team

Thirteen AI agents, each with a defined role, a name, and a private memory that improves over time. They work in Slack — message any of them directly, or they'll come to you when something needs your attention.

Loading team…

How the team communicates

Every agent has its own Slack channel. You talk to them like you'd message a colleague — they respond, take action, and follow up. Agents can also message each other internally when a task needs a different specialist. Ahsoka (performance coach) runs weekly check-ins with every agent via internal messaging — you don't see the chatter, just the outcomes.

How agents get better over time

You give feedback

Tell an agent it got something wrong, or confirm something worked well. That correction gets written into the agent's permanent memory immediately.

Ahsoka runs weekly reviews

Every agent gets a 1:1 coaching session. Ahsoka checks if they're applying previous feedback, identifies what's working, and writes specific learnings into their file.

The agent applies it next session

Agents start every session fresh but with their learnings file loaded. The feedback you gave six weeks ago is still in effect — and Ahsoka checks if they're actually using it.

Patterns get escalated

If an agent keeps making the same mistake after multiple coaching sessions, Ahsoka escalates to Sean instead of writing the same note again. Persistent issues get fixed at the system level.

Capabilities

50 skills agents can deploy on demand, plus 12 backend services powering the whole operation. Skills are like specialist tools — agents load the right one for the job, use it, and put it back.

Skills — what agents can do

These aren't just features — they're complete workflows. Each skill contains a full set of instructions, decision rules, and output formats. When an agent runs /brand, it doesn't just read brand colours — it gets the complete brand voice, all naming rules, every do/don't, and full visual identity reference.

Loading capabilities…

Backend services — what runs the platform

These are the 12 Node.js services running on the server. They're the engine room — agents talk to them via API to get data and take action. Each has its own database, its own secrets, and is managed by PM2 (the process manager that keeps them alive).

Loading services…

How It Works

The plain-English version of what's happening under the hood — for Drew, mostly. Sean already knows this bit.

A day in Mando's life

7am

Systems come online

Rex runs the first health check. All 12 services get pinged. If anything's wrong — a crashed process, an expired token, an error spike from overnight — Rex flags it in #operations-lead before anyone in the office arrives.

8am

Cara and Fennec brief Sean and Drew

Each EA checks overnight emails, reviews the calendar for the day, and prepares a briefing. If there's an urgent email or a call coming up in the next 2 hours, they'll send a heads-up without being asked.

9am

Boba starts working the lead pipeline

30 new leads get created in Pipedrive. Each one is researched across Serper, Apollo, ABN lookup, and website scraping — and arrives with a full profile and an outreach draft ready to send.

All day

Agents respond in real time

Any Slack message to an agent channel gets a response. Ask Elara about ad performance. Ask Rex why a service is down. Ask Cara to draft an email. Ask Kira to research a candidate. They're always on.

8pm

Luca publishes content

The daily social content generator runs. Posts for Facebook, Instagram, LinkedIn, and Google Business Profile go through Elara's review, get approved, and are queued for publishing — automatically.

Every 2hrs

Heartbeat checks

Every 2 hours from 7am to 10pm, a lightweight check runs across the whole system. It only sends a Slack message if something changed — if everything's fine, it stays quiet. No noise.

Sun

Weekly planning and reviews

The social content calendar gets planned for the week ahead. Ahsoka runs performance reviews with every agent. The learning loop closes — what worked this week gets reinforced, what didn't gets corrected.

How Mando remembers things

Agents don't have unlimited memory — each session starts fresh. But they're designed around a 5-tier system that means they always have the right context for the job.

Your direct feedback Highest priority

If you tell an agent something is wrong or works well, it gets written into that agent's permanent learnings file immediately. This is the only memory that can never be overridden — your feedback always wins.

Always-loaded knowledge

A set of core files — who each agent is, the rules they follow, team-wide knowledge, IC brand and client context — are loaded automatically at the start of every session. Think of it as the briefing file they read before starting work each day.

Recent events

A rolling 7-day log of anything business-significant — decisions made, campaigns paused, issues resolved. Agents add to this and read it so the whole team stays in context, even across separate conversations.

Smart search (on demand)

16MB of conversation history and client knowledge, searchable in seconds. When an agent needs to recall something specific — "what did Drew decide about the Gold Coast campaign?" — they search for it rather than trying to load everything upfront.

Per-client knowledge files

Dedicated files for ECTC, Australis Roofing, and Increase Construction itself. Campaign history, key contacts, decisions, and context — stored separately so knowledge about one client never bleeds into conversations about another.

What happens when you give feedback

This is the most important thing to understand about the system. Every correction you give is permanent. If you tell Elara that she approved copy that sounded too corporate, that rule goes into her memory file and applies to every review she does from that moment on — forever, unless you change it. The more feedback you give, the better the team gets. It compounds.

Brand Style Guide

The complete visual and verbal identity for Increase Construction. This is the source of truth Mando uses for all content. Reviewed by Elara.

Logo

The wordmark is the logo — there is no separate icon. The "IC" monogram is for favicon/app-icon contexts only where the wordmark won't fit.

White version on dark/navy backgrounds. Colour version on light backgrounds.
Minimum digital size: 120px wide. Clear space equal to the cap height of "I" on all sides.
Never stretch, recolour, rotate, add effects, or place on busy photographic backgrounds.
Never use the IC monogram as a standalone logo on documents or marketing materials.

Colours

Deep Navy

#1a2744

Steel Blue

#225985

Gold

#cf8e25

Footer Dark

#161e2a

Header Text

#45505d

Muted

#7f8ea1

Light BG

#f5f7fa

Success

#16a34a

Error

#dc2626

Typography

Poppins for all headings (600–700 weight). Open Sans for all body copy (400–600 weight). Minimum 16px body size. No exceptions.

H1 · Poppins 700
36–48px · -0.03em

Win More Work.

H2 · Poppins 700
28–32px · -0.02em

Build Better Teams.

H3 · Poppins 700
20–24px · -0.01em

Grow Without Chaos.

H4 · Poppins 600
15–18px · 0em

The M.O.S.T. Framework

Body lead · Open Sans 400
18–20px · 1.7 lh

We help building and construction businesses win more work, build better teams, and grow without chaos.

Body · Open Sans 400
16px min · 1.6 lh

Most builders come to us with one of three problems: not enough work, not enough of the right work, or not enough team to do the work.

Voice & Tone

Direct, no-nonsense Australian trades talk. Conversational but authoritative. No corporate jargon, no passive voice, no fabricated stats. Em dashes are banned — they're the #1 AI writing tell.

✓ Say this

"Stop hiring in a panic."
"The phone's not ringing — here's why."
"Win more work. Build better teams."
Tie CTAs to revenue impact
Australian English always

✗ Not this

"Leverage integrated workforce solutions."
"In today's competitive landscape…"
"Done For You" as a product label
Em dashes anywhere in copy
"Great question!" as an opener

Critical naming rules

Always	Never
Increase Construction	"Increase Constructions" (no S ever)
BuildAbility	"Build Ability" / "Buildability" / any other variation
Drew Grosskreutz	Any other spelling of the surname
@increaseconstruction.com	@increaseconstruction.com.au — does not exist
Coach / Consultant (Drew's public title)	Calling Drew a "tradie"
National audience	Framing IC as single-state only

Messenger Intelligence System

AI-powered auto-responders for the ECTC, Increase Construction, and Australis Roofing Facebook Pages. Trained on 3,800+ real conversations. Responds in seconds, routes to forms, qualifies candidates, and escalates to humans when needed.

Active Channels

3,887

Conversations Indexed

24,729

Messages Indexed

52/52

Test Scenarios Passed

Live counts as of 12 Jun 2026 — ECTC 1,621 · IC 1,830 · Australis Roofing 436 conversations.

Architecture

The system lives inside social-ads-platform (port 3003). A single webhook endpoint receives events from all subscribed Facebook Pages. The handler identifies which page/channel the message belongs to, loads the correct system prompt, and processes it through a 15-step pipeline.

Facebook Ad

→

Applicant Messages

→

Meta Webhook

→

Caddy Proxy

→

social-ads :3003

→

Claude Sonnet 4.6

→

Reply via Messenger API

Three Separate Systems

ECTC, IC, and Australis Roofing are completely independent. Different prompts, different decision logic, different kill switches, different ad funnels. They share the same handler infrastructure but never cross-contaminate.

Property	ECTC	IC	Australis Roofing
Channel	fb_ectc_page	fb_ic_page	fb_ar_page
Kill switch	MESSENGER_AUTO_REPLY_ENABLED	IC_MESSENGER_AUTO_REPLY_ENABLED	AR_MESSENGER_AUTO_REPLY_ENABLED
Voice	Sean (marketing manager)	Drew (strategic founder)	Australis Roofing recruiter
Primary funnel	TC recruitment (employment forms)	Contractor growth (/get-work, /grow) + recruitment (resume to drew@)	Metal roofer recruitment
Conversations	1,621 (10,554 messages)	1,830 (10,770 messages)	436 (3,405 messages)
Escalation	Slack recruitment channel (tagged ECTC)	Slack recruitment channel (tagged IC)	Handoff to Peter (phone + email required)
Lead router	Yes (form → Messenger confirmation)	No (uses /get-work and resume email)	No

Adset-Aware Routing (IC)

When someone messages from a Facebook ad, Meta sends an ad_id. The handler resolves this to the adset name via the Graph API (cached). The adset name drives the entire response strategy. New adsets are handled automatically at runtime.

Adset Pattern	Funnel	Goal	Key Response
Get Work - [Location]	Contractor	/get-work form	"Our builders need quality [trade] contractors"
Get Work - Roofers [Location]	Contractor (roofing)	/get-work form	Already know the trade, ask crew size
Grow Without Chaos	Business growth	/grow form	"This is exactly who we want to talk to." Escalate if 10+ crew.
Site Supervisor - [Location]	Recruitment (FT)	Resume to drew@	Qualify for insurance exp. No exp = polite decline.
Site Supervisor (GOTT) - [Location]	Recruitment (FT)	Resume to drew@	GOTT-specific roles, same qualification
Estimators/Assessors - [Location]	Recruitment (FT)	Resume to drew@	Assessment/insurance background required
Supervising But Going Nowhere	Recruitment (FT)	Resume to drew@	Skip qualification, pitch the upgrade directly
Roofers Wanted	Ambiguous	Clarify first	"Running your own crew or looking for a full time role?"

ECTC Decision Logic

Action	When	Response
employment_form_qld	QLD applicant with contact details	QLD employment form link
employment_form_nsw	NSW applicant with contact details	NSW employment form link
ask_tc_licence	First contact, no details yet	"Could you shoot me your mobile and email?"
waiting_on_course	Doesn't have TC licence yet	"Reach out once you've got your ticket"
course_redirect	Asks about TC courses	Training page link
pay_rate	Asks about pay	"National award rate, approx $35-36 p/h"
form_completed	Filled out the form	"I'll get your info to the local team"
escalate	Anger, legal, complaints, 5+ msgs	Silent handoff, Slack alert
no_action	Spam, acks ("thanks", "ok")	No reply sent

Lead router integration (ECTC only): When an applicant fills out the employment form, the lead router processes it and calls back to the Messenger system. If the applicant has an open conversation, the bot sends "thanks [name], I'll get your info over to the local area manager" and marks it completed.

IC Decision Logic

Action	When	Response
Contractor Funnel (Get Work Ads)
ask_trade_info	Vague interest, no detail	"What trade, how big's your crew, room for more work?"
get_work_form	Good fit with trade + detail	/get-work link + "I'll text you when I see it"
clarify_intent	Unclear if contractor or job seeker	"Running your own crew or looking for full time?"
grow_form	Fully booked / wants to scale	/grow link + "good problem to have"
Recruitment Funnel (Supervisor/Assessor Ads, Full Time)
ask_experience	First contact on recruitment ad	"Have you had insurance construction experience?"
request_resume	Has insurance exp / good fit	"Shoot your resume to drew@... I'll take a look"
not_a_fit	No insurance exp / irrelevant background	"For these roles we need insurance construction experience"
Business Growth (Grow Ads)
grow_form	Business owner wants to scale	/grow link
escalate	10+ crew (high-value lead)	Silent handoff with context for Drew
Shared
escalate	Wants call, pay discussion, angry, legal	Silent handoff, Slack notification
no_action	Spam, acks, "ok thanks"	No reply

15-Step Handler Pipeline

Every inbound message passes through this pipeline sequentially. Early exits at each checkpoint prevent unnecessary processing.

Skip Echoes

Ignore messages sent by the page itself (prevents reply loops)

Build Referral + Resolve Adset

If from an ad, resolve ad_id to adset name, campaign name via Graph API. Cached in-memory.

Skip Ad Postbacks

Facebook already sends a welcome message for ad clicks. Don't duplicate it.

Dedup

UNIQUE on message ID. Failed/pending messages pass through for auto-retry.

Kill Switch

Per-channel check. When off, messages logged but no reply sent.

Page Token Lookup

Resolves the correct page access token from channels table.

Upsert Conversation

Create/update conversation record with state, metadata, counts.

Escalation Check

If previously escalated to a human, bot stays silent.

Concurrency Guard

Atomic claim prevents parallel LLM calls for the same person.

Profile Fetch

Gets applicant name from Graph API (first contact only, non-blocking).

On-Demand History Fetch

For unknown contacts, pulls conversation history from Graph API before LLM decides.

Lead Router Cross-Reference

ECTC only: checks lead router DB for prior form submissions by phone/email.

Attachment Handling

Images via vision. Stickers: no_action. Voice: "type that out". Files: "email it".

Build Context + Call Claude

Last 10 messages + metadata + adset context. Filters system messages. Resilient parser handles non-strict JSON.

Execute Action

Send reply, escalate, or stay silent. Update state. Process queued messages (depth-limited).

Safety & Resilience

Layer	Mechanism	Purpose
Kill Switches	Per-channel env vars in Infisical	Instant disable without deployment
HMAC Verification	SHA-256 on every webhook	Prevents spoofed requests
Message Dedup	UNIQUE constraint on mid	No double-replies
Concurrency Guard	Atomic claim per PSID	Prevents parallel LLM races
Silent Handoff	Bot stops + Slack alert	Applicant doesn't know it was a bot
24hr Window	Enforced on all outbound paths	Meta messaging policy compliance
Auto-Retry	6-hourly cron sweep	Self-healing, no messages lost
Resilient Parser	Extracts JSON from LLM reasoning	Handles non-strict output
System Msg Filter	3-tier: storage, context, prompt	Facebook internals never reach LLM
Recursion Limit	Depth 2 for queued messages	Prevents runaway LLM calls

Hard Rules (Both Channels)

NEVER confirm work available in a specific location
NEVER promise start dates, shifts, or hours
NEVER discuss pay rates (escalate if pushed)
NEVER say "staffing company" or "labour hire"

NEVER use em dashes in generated content
NEVER ask for info already provided
NEVER send the same message twice
NEVER say IC "has work" (the builders have it)

Background Tasks (Every 6 Hours)

Follow-Up Nudge

Form sent 4+ hours ago, no reply, still within 24hr window. One nudge per conversation.

Message Retry

Sweeps failed/pending inbound messages. Reprocesses through full pipeline.

Stale Marking

24hr window expired conversations marked stale.

Environment

Variable	Purpose
FB_APP_ID / FB_APP_SECRET	Meta App credentials + HMAC verification
FB_WEBHOOK_VERIFY_TOKEN	Webhook subscription verification
MESSENGER_AUTO_REPLY_ENABLED	ECTC kill switch
IC_MESSENGER_AUTO_REPLY_ENABLED	IC kill switch
OPENROUTER_API_KEY	Claude Sonnet 4.6 via OpenRouter
SLACK_BOT_TOKEN	Escalation notifications

Key Files

File	Purpose
lib/messenger/handler.js	15-step pipeline, LLM calls, retry logic, stale check
lib/messenger/system-prompt.js	ECTC + IC prompts, channel registry
lib/messenger/meta-api.js	Graph API: send, profile, typing, ad context resolver
server.js	Webhook routes, admin routes, form callback

Database

Table	Purpose	Key Columns
messenger_conversations	Per-applicant state	psid (PK), channel_id, state, first_name, last_action, metadata (adset, referral, lead_router)
messenger_messages	Audit trail + dedup	mid (UNIQUE), direction, message_text, ai_action, processing_status
messenger_escalations	Escalation tracking	psid, reason, slack_ts, resolved

Cost

~$5-15/month per channel via OpenRouter (Sonnet 4.6). Each message ~$0.002-0.005 in tokens. Infrastructure: $0 (existing VPS).

Monitoring

Live dashboard in the ECTC Leads → Messenger tab. Stats, searchable conversation table, chat-bubble viewer, escalate/reset actions. Proxied via /api/proxy/messenger/*.

Technical Reference

For Sean. The complete infrastructure, auth, file paths, secrets, and operational runbooks.

Auth headers by service

Loading…

Secrets — Infisical (primary)

Self-hosted at 127.0.0.1:8443. 80 secrets across 13 scoped folders. Machine identity mando-bootstrap (Universal Auth) authenticates start-mando.sh at boot. Infisical is the sole secrets store — SOPS was decommissioned 2026-04-08; there is no SOPS/Age fallback. If Infisical auth fails or returns fewer than 70 keys, start-mando.sh aborts the boot (fail-closed).

🚨 Everything in Claude Code conversations goes to Anthropic's API. Never run pm2 env, cat .env, or printenv. A PreToolUse hook blocks this automatically.

Infisical Folder	Contents
/shared	MC_INTERNAL_KEY, cross-service service account paths
/clients/ic	Increase Construction client credentials
/clients/ectc	East Coast Traffic Control credentials
/services/mission-control	MC_ADMIN_TOKEN, MC_READONLY_TOKEN, OPENROUTER_API_KEY, SLACK_BOT_TOKEN…
/services/social-ads-platform	Facebook/LinkedIn/Twitter/GBP OAuth tokens
/services/meta-ads-mcp	META_ACCESS_TOKEN, ad account IDs
/services/google-ads-mcp	Google Ads OAuth tokens, developer token
/services/luca-platform	LUCA_AUTH_TOKEN, OpenAI/Gemini keys
/services/meeting-summarizer	Gemini API key, Zoom credentials
/services/bdm-bot	Serper, Pipedrive, Apollo keys
/services/ectc-lead-router	WEBHOOK_SECRET, OpenAI key, email credentials
/services/calendar-reminder	DREW_EMAIL, JUSTCALL_* — consumed by MC routes/calendar.js (service retired, folder live)
/services/video-content-processor	Drive folder ID, Gemini key

Key file paths

/home/node/.openclaw/openclaw.json

Production OpenClaw config — owned by ubuntu:ubuntu (Docker uid 1000). Always edit this one.

/home/sean/.openclaw/openclaw.json

Dev config — completely separate. Changes here don't affect production.

/home/node/.openclaw/agents/{id}/agent/

Per-agent files: AGENT.md, SOUL.md, IDENTITY.md, USER.md, learnings.md

/home/node/.openclaw/workspace/

Shared workspace bootstrapped into every session: MEMORY.md, SOUL.md, AGENTS.md, HEARTBEAT.md, TOOLS.md

/home/node/.openclaw/workspace/memory/semantic/

Per-client knowledge: ectc.md, australis-roofing.md, increase-construction.md

/home/sean/.claude/skills/

50 Claude Code skills — one SKILL.md per directory. Add a dir with SKILL.md and it's auto-discovered.

MANDO_WORKSPACE_SA_JSON (Infisical /shared)

Google service account env var -- shared by all services. Impersonates drew@, sends from mando@increaseconstruction.com

/home/sean/openclaw/start-mando.sh

Canonical restart — always use this, never pm2 restart (Infisical won't inject)

/home/sean/infisical/docker-compose.yml

Infisical Docker stack (app + postgres + redis) at 127.0.0.1:8443

Ownership rules

⚠️ Files under /home/node/.openclaw/ must be owned by ubuntu:ubuntu (Docker uid 1000). When sean edits files there, ownership drifts and Docker can't read them.

# Fix single file
sudo chown ubuntu:ubuntu /home/node/.openclaw/openclaw.json && sudo chmod 660 /home/node/.openclaw/openclaw.json

# Fix everything after bulk changes
sudo chown -R ubuntu:ubuntu /home/node/.openclaw/

Adding a secret safely

# Paste value at prompt — never inline in the command
infisical secrets set SECRET_NAME \
  --projectId 616a8bc2-e838-4d95-8ea8-a78ebeee889f \
  --env prod --path /services/mission-control

# Verify it exists (name only — never print the value)
infisical secrets get SECRET_NAME --projectId 616a8bc2... --env prod \
  | grep -c SECRET_NAME  # returns 1 if set

GitHub repos

Directory	Repo	Note
mission-control/	increase-construction/mission-control
social-ads-platform/	increase-construction/social-ads-platform
meta-ads-mcp/	increase-construction/meta-ads-mcp
google-ads-mcp/	increase-construction/google-ads-mcp
luca-platform/	increase-construction/luca-platform
meeting-summarizer/	increase-construction/meeting-summarizer
bdm-bot/	increase-construction/bdm-bot
seo/	increase-construction/seo
calendar-reminder/	increase-construction/calendar-reminder
video-content-processor/	increase-construction/video-content-processor
ectc-lead-router/	increase-construction/ectc-lead-router
pipedrive-scorer/	increase-construction/pipedrive-scorer
increaseconstg/	increase-construction/increaseconstruction-website	⚠️ Dual remote — push to BOTH origin (WP Engine) AND github
~/.claude/skills/	increase-construction/mando-skills

Alerts