Raju Rayhan
Head of Development · AI Platform Architect
Available · Q3 2026
← work / the hive
Case 03 · 2024 · Internal AI operations

The Hive

An internal operations platform for Lighthouse Graphics. Meeting summaries, action items, and project scopes generated automatically. Yelp-inbound business leads answered autonomously by an AI agent. Slack, ClickUp, and GHL woven together so information stops falling between people.
LaravelNestJSReactOpenAISlackClickUpGHLYelp APIn8n
1d
Avg. lead response → min
4
Systems stitched together
0
Action items lost to memory
Σ
Auto-summarized meetings
§ 01 The brief
Context

A services agency runs on small, constant acts of coordination — recapping meetings, turning talk into tickets, answering inbound leads before they cool. The Hive was built to do those acts automatically so the humans could stay in the interesting parts of the job.

Constraints
  • Had to slot into the existing Slack / ClickUp / GHL stack — no new surfaces for the team.
  • Summaries had to be trustworthy enough that people stopped taking their own notes.
  • Yelp-inbound leads needed human-quality replies within minutes, any hour.
  • Everything auditable — who said what, what the AI said, what it did next.
The Hive shouldn't feel like a bot. It should feel like a diligent junior who happens to never sleep, never forget, and always writes things down.
— product principle, pinned
§ 02 The architecture simplified
   ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐
   │  Slack   │  │ ClickUp  │  │   GHL    │  │   Yelp   │
   │ webhooks │  │  tasks   │  │   CRM    │  │ webhooks │
   └────┬─────┘  └────┬─────┘  └────┬─────┘  └────┬─────┘
        │             │             │             │
        └─────────────┴──────┬──────┴─────────────┘
                             │
                    ┌────────▼─────────┐
                    │   NestJS gateway │   ← event bus · retries
                    └────────┬─────────┘
                             │
               ┌─────────────┼─────────────┐
               │             │             │
      ┌────────▼──┐  ┌───────▼────┐  ┌─────▼─────────┐
      │ Meeting   │  │ Lead       │  │ Scope         │
      │ summarizer│  │ responder  │  │ generator     │
      │ (OpenAI)  │  │ (OpenAI)   │  │ (OpenAI)      │
      └────────┬──┘  └───────┬────┘  └─────┬─────────┘
               │             │             │
               └─────────────┼─────────────┘
                             │
                    ┌────────▼─────────┐
                    │  Laravel core    │
                    │  · audit log     │
                    │  · humans in UI  │
                    └──────────────────┘
§ 03 A few decisions
03.1
Agents own an outcome, not a feature

Rather than one monolithic AI brain, The Hive is a small fleet of focused agents — one writes meeting summaries, one handles Yelp leads, one drafts scope documents. Each has narrow permissions, its own prompts, its own evaluation set. When something regresses, we know exactly which agent to fix.

03.2
Humans stay in the loop on anything outbound

Internal artifacts (summaries, action items, tickets) post automatically. Anything going to a customer — even a Yelp reply — is drafted, routed into Slack for a one-click approve, then sent. Trust grows before autonomy does.

03.3
One write path to every tool, never two

Every system-of-record write goes through the Laravel core. Agents don't touch ClickUp or GHL directly — they hand off structured intent, the core validates and writes. The audit log tells you who asked, who approved, who executed.

03.4
Prompts live in the repo, versioned like code

Prompts aren't config — they're program text. They live in the monorepo, reviewed in PRs, with evaluation runs on change. 'Fixing the prompt' is a commit, not a Slack message someone forgot about.

§ 04 Outcomes
Before
  • Meeting notes — when they existed — lived in someone's Notion
  • Action items surfaced in standup two days late
  • Yelp leads answered whenever someone happened to check
  • Project scopes drafted from scratch, every time
After
  • Summaries and action items in Slack minutes after a call ends
  • Tasks auto-created in ClickUp, assigned by heuristics + review
  • Yelp leads answered within minutes, approved by a human first
  • Scope drafts generated from call transcripts, ready to edit
Next case
AalapAI — WhatsApp × AI
Read next →