Most agencies sell features. We sell a knowledge graph, an agent fleet, and a structure that compounds. The Colliers OSBA engagement would land on this foundation — not as a one-off build, but as a tenant on a system we run end-to-end for life-sciences commercial operations.
Stages 1 & 2 are what you'd pay for first. Everything underneath this page is what makes Stages 3 through 7 cheap to add later — because the foundation is already there. New capabilities are weeks, not months.
The foundation is a multi-tenant Postgres knowledge graph (we call it Hub DB), with per-client schema isolation and embeddings on every entity. It's the same architecture currently in production for multiple life-sci clients. Nothing on this site — the receipts, the corpus inventory, the action items, the funnel framing — would exist without it.
Bay Area life-sci tenant universe + signals + relationships + past deals + past conversations — bound to the company, not the rep's inbox. Persistent. Versioned. Embedded for semantic retrieval.
Pre-meeting briefs, transcripts, post-call recaps, decks, LOIs, redlined contracts — bound to the opportunity record. Reps don't lose context when a deal goes quiet for three months.
Every signal we ingest is timestamped, attributed, weighted, and linked to the account(s) it touches. Decays over time. Re-weighted when other signals corroborate.
Fireflies / Fathom transcripts → Hub DB with speaker segmentation, intent flags (task, pricing, metric, question), and embedding-indexed chunks. Action items auto-extracted by speaker. Already done for the Colliers 5/22 + 5/28 transcripts.
Per-tenant data isolation isn't a feature we'd add — it's the default. Your Champions DB, your conversation history, your signals never touch another client's tenant. Encrypted at rest. Exportable at handoff. The corpus already sitting in Hub DB for Colliers (25 research files, 2 transcripts, 13 action items) demonstrates the model.
BC runs as an agentic org — not a chatbot, not a single "AI assistant." Seven specialist directors coordinate work the way a human executive team would. Each director has a fleet of sub-agents that execute discrete tasks under an Anthropic orchestrator-worker contract (goal · context · tools · done-condition). A CEO Conductor sits above all directors as the human-gated dispatcher.
Dev · Sales · Operations · Marketing · R&D · Legal · Finance — orchestrated by the CEO Conductor.
News, social, clinical trials, NIH grants, patents, SEC filings — extensible per client.
Brain (KB) for reasoning + RAG. Heart (Ops) for orchestration + workflows. Independent scaling.
Full IC. Writes, reviews, maintains all confirmed code. Builds sub-agents for other directors.
LivePipeline mgmt, targeting/scoring, proposals, MI, Miller-Heiman, blue sheets. The director that'd own Colliers' Stages 1–3 day-to-day.
PartialOutbound content, brand, social, paid ads, event mgmt, multi-touch. The director behind the marketing stack.
LivePost-signed client work, FRO support, intelligence-pipeline orchestration, catch-all coordinator.
DesignedLearning library, market scanning, evergreen idea queue, prototyping. Where new capabilities incubate.
DesignedPost-signature contract wrap, redline reviews, compliance tracking. Walled off, Hub-DB-only access. Powers the Stage 4 legal agent.
DesignedCash ledger, invoicing, AR/AP, P&L, tech-stack cost tracking. Walled off, Core-DB-only access.
DesignedSits above all directors. Plans across them, dispatches work, holds the human-gated Cortex Inbox where Brian confirms every proposal.
LiveEvery director → sub-agent dispatch specifies: (1) Goal — what the sub-agent achieves, one sentence · (2) Context — what it needs to know that it doesn't already · (3) Tools — explicit tool allowlist · (4) Done condition — verifiable, not subjective.
If any of the four is missing, the sub-agent drifts. This is enforced in every director's system prompt. It's why work doesn't quietly fail.
Your Stages 1 & 2 work would run under the Sales Director. If you add the marketing stack, that's the Marketing Director's fleet. If you add the legal agent, that's Legal Director sub-agents. Three different domains, one orchestration layer, one knowledge graph, one place where work proposes itself for human approval. You don't manage seven different vendors — you have one BC partner running seven coordinated specialists.
BC's infrastructure is split across two VPS hosts — not for redundancy, for clean separation of concerns. Brain (KB VPS) is where LLM reasoning, RAG retrieval, and council-style multi-model decisions live. Heart (Ops VPS) is the pump — n8n workflows, heartbeats, scheduled jobs, Hub DB writes, integrations. No LLM lives on Heart; no scheduling lives on Brain. Each does one job well.
Single-model chat (claude-sonnet, gpt-4o, gemini, qwen-local for cheap classification). Multi-model council for high-stakes decisions (3 models fan out, judge synthesizes, all audit-logged). RAG over Brain DB — research_reports, council_decisions, kb_chunks, drug_candidates, companies, pubmed_articles, news_articles, patents — ~30 tables of pre-indexed life-sci context.
n8n workflows fired by webhook or schedule — Gmail digests, Fireflies ingestion, social listeners, signal scrapers, content publishers. Hub DB read/write across target_dossiers, new_accounts, contacts, opportunities, plus your per-tenant Colliers schema. Heartbeat log for every workflow run — observable, debuggable, no silent failures.
When a lease-expiration signal fires at 6:47 AM, Heart catches it, writes it to your tenant schema, and triggers the matcher workflow. Heart then asks Brain "score this signal against this account's context"; Brain returns a routed action with the right rep, the right Gold Sheet attached, the right first-message draft. By the time Kareem's reps log in at 7:30 AM, the daily list is waiting. Two systems, one outcome, fully observable end-to-end.
The platform isn't a feature. It's the substrate every feature compounds on top of — same graph, same fleet, same harness, every client.
Every BC agent — signal scraper, content drafter, transcript ingester, prospect researcher, redline checker — is built on the same reusable harness. New capabilities are configuration, not architecture.
News intel, social listening, clinical-trial collection, NIH grant matching, patent indexing, SEC filings, PubMed/OpenAlex research, account enrichment, ICP scoring, message generation, transcript ingestion. Each connects to a public or commercial API, runs through Python or Colab, stores results in Hub DB. Every skill feeds into or receives from at least one other skill — compounding intelligence.
Every agent has explicit, allowlisted access to the same MCP (Model Context Protocol) tools the human team uses. Same tools, same audit trail. Agents don't have hidden capabilities; humans don't have to learn agent-specific surfaces.
Hard rule across all agents: write actions surface as proposals in the Cortex Inbox; nothing executes until Brian (or the designated human gate — for Colliers, that'd be your team) confirms. Eliminates "agent went rogue" failure modes. The price is one extra approval step; the value is total auditability.
Once the foundation is in — knowledge graph, director fleet, Brain×Heart back-end, agent harness — new capabilities stack on top instead of being rebuilt from scratch. That's why we can quote Stages 3, 4, and 7 as natural extensions rather than separate projects. It's also why each new capability gets cheaper, not more expensive, over time.
End-to-end commercial platform for an industrial centrifuge manufacturer — account graph, ICP scoring, daily intelligence brief, campaign ops, EmailBison integration. Same architecture as what we'd build for Colliers, different vertical.
Cell-and-gene-therapy CDMO partnership with full MI pipeline, content studio, brand intelligence, partnership tracking. Same Brain × Heart split. Different sub-agent fleet.
BC's fictional Series-A — full demonstration of the stack: brand → site → outbound → scoring → SEO → pipeline. Public; visit to see what an integrated build looks like.
Here's the translation back to your funnel. The core engagement — what Kareem asked for on May 14 and refined on May 28 — uses Layers 1–4 to deliver Stages 1+2. Everything else on this site (Stages 3, 4, 7 + marketing) is the same foundation, pointed at different problems.
| Layer | Used for | Today's role at Colliers | Tomorrow's role |
|---|---|---|---|
| Knowledge graph | Hub DB | Account · signal · conversation graph for Colliers tenant | Stage 3 Gold Sheets · Stage 4 contract memory · Stage 7 expansion history |
| Agent fleet | Director-tier specialists | Sales Director runs the matcher and message draft | Marketing Director adds content; Legal Director adds redlines |
| Brain × Heart | VPS infrastructure | Daily 6 AM signal sweep, scoring, routing | Real-time signal listener, multi-channel campaign exec, contract-review pipeline |
| Agent harness | Reusable scaffold | 3 sub-agents (scorer · matcher · message drafter) | +5 more for Stages 3–7 in subsequent rounds |
| Cortex Inbox | Human approval gate | Kareem (or designate) confirms daily routing decisions | Same surface for marketing publishing, legal redlines, expansion alerts |
This isn't aspirational architecture; it's the same stack we run our own business on, and the same stack we deploy for life-sci clients today. Colliers would be the first CRE deployment — the architecture is industry-agnostic, the signal stack is what we'd customize.
