Sticks App Guide

Sticks is a personal AI assistant powered by Claude. It provides real-time insights on LFIQ properties, market data, and internal knowledge, accessible via chat.

What It Does

Sticks is an AI assistant tailored to LFIQ knowledge:

Property queries: "What's the occupancy at 123 Main?" → Returns current property data
Market analysis: "What are rents in the Mission?" → Analyzes market trends
Deal analysis: "Should we bid on this property?" → Claude analyzes comps, cap rate, market risk
Knowledge retrieval: "What did we learn about XYZ?" → Searches internal documents, observations, research
Recommendations: Claude suggests actions based on portfolio data and market conditions

Primary features:

Chat interface (web and iMessage)
Context-aware responses (access to portfolio data, market data, history)
Document search (Box, SharePoint, internal documents)
Property briefing (current metrics, leases, rent trends)
Multi-turn conversations (thread memory across sessions)

Deployment

| Environment | URL | Status | Platform | |-------------|-----|--------|----------| | Production | https://sticks.lfiq.app | Live | Vercel + Cloud Run | | Preview | https://sticks-branch.lfiq.app | Auto-deploy on PR | Vercel | | Local Dev | http://localhost:3006 | Via npm run dev | Local machine | | Backend Chat Proxy | Cloud Run | Processes queries, validates OIDC | Cloud Run |

Tech Stack

| Component | Tech | Notes | |-----------|------|-------| | Frontend | Next.js 15 | React 19, chat UI, thread management | | Language | TypeScript | Full type coverage | | Auth | Clerk | OAuth via accounts.lfiq.app | | Backend | Cloud Run | Python FastAPI, prompt engineering | | AI Model | Anthropic Claude | Latest available model (claude-opus or claude-sonnet) | | Context | Neon + Pinecone | Portfolio data, vector search for RAG | | Deployment | Vercel (frontend) + Cloud Run (backend) | Auto-deploy on main |

Local Development

Start the App

cd /path/to/02-brick.apps/apps/sticks
npm run dev
# Runs on http://localhost:3006

Environment Variables

| Variable | Required? | Purpose | |----------|-----------|---------| | NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY | Yes | Clerk OAuth | | CLERK_SECRET_KEY | Yes | Session signing | | NEXT_PUBLIC_API_URL | Yes | Backend API base (Cloud Run) | | ANTHROPIC_API_KEY | Yes | Claude API key (backend only) | | HUB_CHAT_PROXY_SA_JSON | Yes | Cloud Run invoker SA (base64) |

Pull from Vercel:

vercel env pull

How It Works

Sticks Architecture

User Chat Input
  ↓
Vercel Frontend (sticks.lfiq.app)
  ↓
Cloud Run Chat Proxy (validates OIDC token)
  ↓
Prompt Assembly:
  1. System prompt (tuned for LFIQ context)
  2. User question
  3. Context (property data, recent observations, market trends)
  4. Document snippets (RAG from Pinecone)
  ↓
Anthropic API (Claude Opus or Sonnet)
  ↓
Response streamed back to frontend
  ↓
User sees answer in chat

System Prompt Design

Sticks' system prompt tells Claude:

You are Sticks, an AI assistant for LFIQ
You have access to portfolio data (properties, rents, leases)
You have access to market data (trends, comps, rent history)
You have access to internal observations and knowledge
Answer in plain operator voice
Provide reasoning, not just answers

Example system prompt:

You are Sticks, an AI assistant for LFIQ. Your job is to help the investment team 
understand their portfolio and make better decisions.

You have access to:
- Portfolio data: properties, units, rents, leases, valuations
- Market data: rent trends, competitor activity, comps, distress indicators
- Internal knowledge: observations from 14 data sources, prior analysis

When asked a question, use this data to answer. Show your reasoning. If data is missing, say so.
Speak like a real estate operator. No jargon. No unnecessary caveats.

Key Flows

Flow 1: Answer Property Question

User types: "What's the occupancy at 123 Main?"
Frontend sends to Cloud Run proxy
Proxy validates Clerk OIDC token
Proxy retrieves from Neon:
- Property record (address, units, units occupied)
- Lease status (upcoming expirations)
- Rent data (current market, achieved rents)
Context assembled: "Property 123 Main has X units, Y occupied (Z%), market rent is $M"
Claude responds: "Occupancy is Z%. Current market rent is $M, you're getting $M+. Lease expirations: [list]"
Response streamed to user

Flow 2: Multi-turn Conversation

User: "Should we increase rent at 123 Main?"
Sticks: "You're currently at $2,000. Market is $2,200. Risk is tenant turnover."
User: "What's tenant turnover rate in the neighborhood?"
Sticks: (retrieves historical turnover data) "Turnover in neighborhood is 15% annually. Your property is 10%. Strong performance."
User: "Show me comparables"
Sticks: (retrieves comps from Pinecone) "Here are 5 similar properties in the area with rent data..."

Flow 3: RAG (Retrieval-Augmented Generation)

User: "What did we learn about the Sunset district?"
Frontend sends to Cloud Run proxy
Proxy queries Pinecone for embeddings related to "Sunset district"
Top 5 documents/observations retrieved
Claude reads context and synthesizes answer
Response includes sources: "Based on observations from [dates], here's what we learned..."

Chat Context

Sticks can access the following data for context:

| Data Type | Source | Purpose | |-----------|--------|---------| | Properties | Neon portfolio schema | Address, units, occupancy, rent roll | | Rent trends | Neon market schema | Historical and current rent data | | Observations | Neon items schema | Intel observations (delinquency, lease expirations, alerts) | | Comparables | PropertyRadar (via Stacks) | Market comps, cap rates, distress scores | | Documents | Pinecone embeddings | Stored research, investment memos, market reports | | Portfolio metrics | Neon portfolio schema | Revenue, expenses, NOI, occupancy trends |

Troubleshooting

Issue 1: "Chat shows 'Error' state"

Symptom: After typing a question, red error appears
Cause: Cloud Run proxy down, Anthropic API key invalid, or OIDC token expired
Fix:

# Check Cloud Run service
gcloud run services describe sticks-chat-proxy --project=brickston-v2

# Check if service is running
curl https://sticks-chat-proxy.lfiq.app/health

# Verify Anthropic API key
echo $ANTHROPIC_API_KEY

# Clear browser cache and re-authenticate
# DevTools > Application > Clear Site Data

Issue 2: "Claude gives generic answer without property data"

Symptom: "What's occupancy?" → "I don't have access to that data"
Cause: Context retrieval failed, property not found, or prompt issue
Fix:

# Verify property exists in Neon
psql -h ep-tiny-lab-akrddwgy.us-west-2.neon.tech \
  -U command neondb \
  -c "SELECT address FROM portfolio.properties WHERE address LIKE '%123 Main%';"

# Check Cloud Run logs for context retrieval errors
gcloud run logs read sticks-chat-proxy --project=brickston-v2 --limit=50

# Verify Pinecone embeddings are populated
# Log in to Pinecone console: https://app.pinecone.io
# Check vector count in index

Issue 3: "Claude reference data that's incorrect or outdated"

Symptom: "What's the rent at 123 Main?" → Returns data from 3 months ago
Cause: Neon data is stale, embeddings not updated, or context window too small
Fix:

# Verify Neon has latest data
psql -h ep-tiny-lab-akrddwgy.us-west-2.neon.tech \
  -U command neondb \
  -c "SELECT created_at, monthly_rent FROM portfolio.units WHERE property_id = $1 ORDER BY created_at DESC LIMIT 1;"

# Re-index Pinecone with latest documents
gcloud run jobs execute sticks-embeddings-refresh --project=brickston-v2

# Check Cloud Run logs
gcloud run logs read sticks-embeddings-refresh --project=brickston-v2

Common Tasks

Task 1: Ask Sticks a Property Question

Open https://sticks.lfiq.app and ask:

"What's occupancy at [property name]?"
"What rents are we getting at [property]?"
"Should we bid on [address]? Compare to comps."
"What are lease expirations at [property] in the next 90 days?"

Task 2: Search Knowledge Base

Ask Sticks:

"What did we learn about [neighborhood]?"
"Show me recent investment memos on [topic]"
"What's our analysis on [property type]?"

Task 3: Analyze Market Trends