Most RAG portfolio projects are a chatbot over one PDF for one user. TenantAgents is built closer to how real SaaS products work, many companies on one platform, each one completely walled off from the others, with an AI agent that does not just talk, it does things.

Building an AI assistant over one company's documents is easy enough. The hard part is doing it for many companies on the same platform, safely: Company A should never be able to see Company B's documents, chat history, or data, not even by accident, not even if there is a bug somewhere in the app.

• Each company gets its own private workspace. Upload documents, chat with an AI that only knows that company's documents, never anyone else's.
• The AI can take real actions, not just chat. Ask it to create a support ticket or send an email, and it decides to do that on its own using Gemini's function calling, not a hardcoded "if user says X, do Y" script, and every action gets logged to an audit trail.
• Every answer shows its work, which source documents it used, how confident the match was, how long it took, and how many tokens it used.
• Isolation is enforced at the database level, not just hidden in the UI. Every table that holds a company's data has a tenant_id column, and every query checks it, so even a bug in the app logic could not leak data across companies.

1. Sign in and pick a company. Two demo companies are ready to go, Acme Corp and Globex Inc, each with its own workspace.
2. Upload documents to the Knowledge Base. A .txt or .md file gets split into chunks, embedded, and indexed, ready to be searched.
3. Ask the Chat Assistant anything about the documents. It answers using only that company's documents, and shows which sources it pulled from and how confident it was.
4. Ask it to do something, like send an email or open a ticket. It decides on its own whether the request needs an action, and every action shows up in the Audit Log with a timestamp and status.
5. Check the Dashboard for a live view of the workspace, documents uploaded, questions answered, actions taken, and recent activity.

Getting a demo running locally is maybe 40% of actually shipping something. A few of the real problems that ate the rest of the time:

• A mystery infinite request loop, only in dev. Turned out OneDrive was silently syncing the project folder, re triggering Next.js's file watcher every second.
• A 404 loading Clerk's own script, only in production. The middleware's path matcher was accidentally excluding the exact path Clerk needs to proxy its script through.
• Redesigned how agent actions work three times. The first version wired actions through n8n to fire real Slack messages. Eventually cut it, an external dependency adding complexity without adding real value over a simple internal audit trail.

A lead fills out a form and then sits in a spreadsheet until someone checks it. Manual triage means hot leads wait, and interest fades fast. This workflow removes that step entirely.

1. Lead submits a form (Google Forms), the response lands in Google Sheets.
2. n8n polls the sheet for new rows.
3. The lead is scored Hot or Warm, and a data quality flag is applied.
4. A budget threshold check routes the lead: $10,000 or more goes to the #sales-leads Slack channel, under $10,000 goes to #nurture-leads.
5. Slack delivery includes error branch handling (Continue On Fail), so a failed send is never silently lost.

• Import lead-routing-automation.json into an n8n instance.
• Connect your own Google Sheets and Slack credentials.
• Update the budget threshold in the "Score & Route Lead" node if needed.

Built as a portfolio demo with test data. The threshold, scoring logic, and notification channels are easily adapted for a real pipeline, for example CRM write back, SMS alerts, or multi tier routing.

In real estate, deals are usually lost because of slow response times rather than a lack of interest. If a hot lead sits untouched for even a few hours, they will likely move on to a competitor. The goal was a reliable workflow where the second a lead engages, moving into the Contacted stage, the entire sales team gets pinged instantly.

• Custom deal pipeline. A seven stage pipeline matching a standard property buying journey: New Lead → Contacted → Qualified → Site Visit Scheduled → Offer Made → Closed Won / Closed Lost.
• Tailored data model. Custom properties across Deals and Contacts to capture Budget Range, Property Type (Apartment, House, Plot, Commercial), Lead Source (Facebook Ad, Referral, Website, Instagram, Walk in), and Site Visit Date.
• Realistic sample data. Fifteen realistic contacts and deals distributed across all pipeline stages, so it reads like a living workspace rather than an empty template.
• Automated lead routing. Whenever a deal changes stage, Zapier catches the update, filters specifically for the Contacted stage, and instantly fires a formatted notification into a dedicated Slack channel with the deal name, budget, property type, and source.

Note on tool limits: HubSpot's free tier does not support native workflow triggers, and Zapier's free plan does not include a native delay action, so the current setup fires instantly on stage change rather than after a delay.

• A three day delay timer with a no activity filter to catch stalled deals.
• Round robin lead auto assignment based on budget and source.
• Custom lead scoring models based on engagement metrics.
• Urgent SMS alerts alongside the Slack notification.

Most portfolio dashboards stop at "here's revenue by month." This one is built the way a RevOps analyst would actually be asked to: not just report what closed, but flag what's stuck, deals sitting in a stage with no movement, quietly turning into lost revenue before anyone notices.

A deal is flagged stalled once it has sat in one stage for more than 30 days with no movement, a standard RevOps threshold for "this needs a human, not a report." The dashboard sorts these worst first and color codes severity, so a manager can open the workbook Monday morning and know exactly which deals to ask about in stand up.

The dataset is a reproducible synthetic dataset with a fixed random seed, not a raw drop in of a public dataset. An off the shelf retail transactions dataset has no CRM stage history, so it cannot support a real funnel or a genuine stalled deal analysis, the two things this project is actually trying to demonstrate.

• $14.77M in closed won revenue across 887 deals, a 22.7% overall win rate.
• Subscription revenue dominates the mix at 88%, with Services and Add on each around 6%.
• Funnel conversion holds steady in the low to mid 70s% at every stage, with the sharpest drop off between Negotiation and Closed Won at 51%.
• East region and rep Jordan Ellis lead on closed revenue.
• $1.18M, 14.9% of open pipeline, is currently stalled, the action item this dashboard exists to surface.