Architecture

Tool Catalog — 9 typed tools

Why This Matters

ENERGY STAR Portfolio Manager holds data on 600,000+ commercial buildings — but its interface requires manual queries and CSV exports. This MCP server lets any Claude-powered workflow access that data through natural language. Ask "What's the EUI trend for my NYC office portfolio?" and get structured data back, ready for analysis. Published open-source for the sustainability community.

RAG Pipeline

Knowledge Base Composition

Interactive Demo

Delivery Interfaces

Analysis Pipeline

Accuracy Targets (vs ASHRAE Level II)

ECM Categories

Data Sources (15 integrated)

ASHRAE Standards 211, 90.1, 105, 100 · DOE Commercial Prototype Building Models · EIA CBECS & RECS · EPA eGRID · EPA ENERGY STAR Portfolio Manager · WaterSense Standards · DOE BPS Policy Tracker · DSIRE Incentives Database · RS Means Construction Cost Data · NREL Cambium Dataset · Engineering firm benchmark data (2025)

Deal Lifecycle

Database Schema

System Architecture

Live Demo — One Record, Every System

A deal entered once fans out across the whole back office. Switch tabs to see each connected capability — the fan-out, the live pipeline dashboard, lease abstraction, and auto-generated marketing.

Verification & Safety

86 automated tests across the platform — the commission math, lease extraction, the agent's tool-use parity against the legacy path, and a security-contract suite that fails closed · a changelog table that records a JSON diff of every record mutation · a read-only database role for the agent (SELECT-only, enforced at the Postgres layer) · OAuth2 PKCE with single-use refresh-token rotation, hardened to volume-mounted persistence after a real production incident.

Second Client — Their Entire System of Record

The same foundation became a second firm's entire system of record. In about two weeks I stood up a 16-module operating platform with its own row-level-security model and 19 database migrations — their deals, contacts, properties, leases, commissions, AR, statements, documents and reporting all live in one place now.

16 frontend modules · 19 migrations · own RLS + explicit-deny policies + activity-trigger lockdown

Why it matters

The first brokerage build proved I could ship a complex operating system. RCB proved I could adapt that operating model to a different brokerage context: different entities, workflows, permissions, reports, imports, and launch constraints. The architecture repeated; the implementation was client-specific.

System surface

Repeatability proof

The point is not that RCB reused a template unchanged. The point is that the same product thesis held up under a second real client: brokerages need one operational backbone for deals, money, documents, reporting, and team workflow.

Tool-Use Loop

Tool Registry — 25 tools across 5 systems

Live Demo — Vera in a Thread

Vera is a Claude tool-use agent living in Slack — a 25-tool registry across five systems (Postgres deal DB, CRM, Notion, Drive, and a RAG corpus of the firm's own documents). It answers plain-English questions in-thread with citations pinned to the exact source, carries persistent memory, and runs scheduled digests (e.g. weekly AR aging). Safety by construction: a read-only database role, an 8-tool-call cap, and every tool call audited.

It Writes Its Own Skills

Vera isn't a fixed prompt. When it works out a reliable way to answer a recurring kind of question — which tables to join for AR aging, how a given firm labels its deal stages, a teammate's standing preferences — it calls skill_write to persist that method as a reusable skill in its own table. On the next invocation those learned skills, per-user preferences and past failures are read back and preloaded into the system prompt. The agent gets more capable the longer a team uses it, without me redeploying anything.

Proactive — Scheduled Digests

“Save this as ‘Monday AR aging’ and run every Monday at 9am” is stored as a saved query; an APScheduler cron inside the container re-runs the full tool-use loop and posts the result to the channel. An optional condition-SQL pre-check skips the run — and the LLM cost — when there's nothing to report. Any teammate can trigger a saved digest on demand.

Safety & Verification

Read-only Postgres role and a SELECT-only guard at the tool layer — writes are rejected even if the guard is bypassed · 8-tool-call cap per question · within-question result cache (a repeated lookup hits the cache, not the API) · retry-once on 429/5xx before counting against the cap · per-tool result truncation · every tool call audited to a JSONB trace. Shipped behind a feature flag with a 15-question parity suite proving the new tool-use path matched the legacy path before cutover.

Live Demo — Lead Conversation

Scope, budget, appointment — handled entirely by the AI. The contractor gets a booked calendar event; the lead gets a confirmation text. Zero human involvement until the site visit.

Lead Response Flow

Admin SMS Interface

Multi-Tenant Architecture

Each contractor is a single database row with their own Twilio number, AI agent name, custom system prompt, Google Calendar connection, admin phones, and service area. New clients require zero code changes — add a contractor record and buy a Twilio number. The AI agent knows the business name, services offered, service area, and operating hours from the contractor's system prompt.

Infrastructure

Multi-tenant on a DigitalOcean VPS · FastAPI on port 8000 · PostgreSQL · Twilio A2P 10DLC (pending carrier approval) · Mailgun inbound email parsing · Claude Sonnet with vision (MMS photo analysis) · Google Calendar OAuth · Nginx + SSL · systemd · Admin SMS commands via Claude intent classification