Offer to close, contingencies, timelines, Wisconsin-specific steps.
- Documents
- 24
- Review
- Quarterly
- Owner
- Tyler Bair
Digital real estate education coach
Real estate questionsshouldn't have to wait.
Leave every person more confident than it found them. Ask anything — AI Tyler answers in plain language, shows the reviewed source behind it, and hands you to Tyler the moment judgment matters more than information.
Conversation with AI Tyler
Educational ModeAI TylerEducation coach · MilwaukeeDesign preview — model not connected
Start anywhere
Ask the question you didn't want to ask out loud.
Every answer is written in plain language, shows where it came from, and tells you what to do next. If it's beyond what I should answer, I'll say so and bring Tyler in.
- Teacher first, never a salesperson
- Plain language before industry language
- Always names the next useful step
BUFF educates. Your licensed representative advises.
AI Tyler is an educational tool. It never advises on negotiation strategy, contingency decisions, appraisal responses, inspection decisions, contracts, transaction management, or deadlines. If you already have a licensed representative, AI Tyler runs in Educational Mode.
Educational use only
BUFF educates. A client's licensed representative advises.
The information provided by BUFF is intended for educational purposes only and should not replace the advice of your own licensed real estate professional, attorney, lender, inspector, or other qualified advisor.
Knowledge architecture
Answers are only as good as what they're built on.
AI Tyler never answers from open-ended model memory. Every response is retrieved from a reviewed, domain-scoped BUFF knowledge base with a named human owner.
Loan types, pre-approval, cash-to-close, rate mechanics. Education only.
- Documents
- 16
- Review
- Monthly
- Owner
- Tyler Bair + lender partner review
Counties, neighborhoods, seasonality, inventory patterns.
- Documents
- 31
- Review
- Monthly
- Owner
- BUFF market desk
Cash flow, duplex strategy, expense modeling, risk framing.
- Documents
- 12
- Review
- Quarterly
- Owner
- Tyler Bair
Buying, Selling and Investor Systems, stages, and client expectations.
- Documents
- 9
- Review
- On change
- Owner
- BUFF ops
Fair housing, licensing, advertising rules, refusal boundaries.
- Documents
- 7
- Review
- Quarterly + on regulation change
- Owner
- Broker compliance
Retrieval pipeline
- 01
Ingest
PDF, DOCX, Markdown, transcript, and web page sources uploaded via admin.
- 02
Normalize
Strip layout, preserve headings, capture source metadata.
- 03
Chunk
Heading-aware chunks (~700 tokens, 12% overlap) keep answers quotable.
- 04
Review
Every chunk requires human approval before it becomes retrievable.
- 05
Embed
Vector index per domain so retrieval can be scoped by intent.
- 06
Retrieve
Hybrid keyword + vector search, top-k with domain boosting.
- 07
Ground
Model may answer only from retrieved chunks; else it escalates.
- 08
Cite
Every claim maps to a source card with title, section, and updated date.
Prompt engineering architecture
Six layers, versioned independently.
The system prompt is composed, not written. Each layer has a different owner and change cadence so compliance language can never be edited by accident while tuning tone.
| Layer | Purpose | Change cadence |
|---|---|---|
| 01 · Identity | Who AI Tyler is, tone, and the education-first promise. | Rarely — versioned with brand. |
| 02 · Boundaries | Hard refusals: legal, tax, lending, valuation, fair-housing risk. | Only with broker compliance approval. |
| 03 · Grounding | Retrieved chunks with source IDs; instruction to answer only from them. | Every request. |
| 04 · Output contract | Answer, plain-language recap, citations, confidence, next step, resources. | Versioned with the UI contract. |
| 05 · Personalization | Journey stage (buy/sell/invest/explore) and prior turns in this thread. | Per conversation. |
| 06 · Escalation | When to stop answering and offer Tyler, with the reason captured. | Tuned from escalation analytics. |
Output contract
Every response must return this shape. The UI renders nothing it cannot verify.
answer — plain language, no jargon without a definitionbullets — optional, max three, each independently usefulcitations[] — title, section, domain, updated dateconfidence — grounded | partial | handing offnextStep — one concrete actionresources[] — matched BUFF assetsescalate — boolean with a captured reason
Hard boundaries
Non-negotiable refusals. Compliance owns this layer.
- Advise on negotiation strategy, contingency decisions, appraisal responses, inspection decisions, contracts, transaction management, or deadlines
- Interfere with an existing agency relationship — BUFF educates, a client's licensed representative advises
- Give legal, tax, lending, or appraisal advice
- Estimate a specific property's value
- Recommend a specific home, offer price, or bid strategy
- Collect sensitive financial details in chat
- Claim to represent the client without a signed agreement
Escalation to Tyler
Knowing when to stop is a feature.
Escalation is not a failure path — it is the product. When any trigger fires, the conversation converts into a prepared handoff: the question, the context, what was already explained, and the specific gap a human needs to close.
- Confidence falls below the grounded threshold
- Question requires legal, tax, lending, or valuation judgment
- User asks about a specific property or writes an offer scenario
- Third clarifying question on the same topic
- User expresses urgency, frustration, or a deadline
- User explicitly asks for a human
Follow-up email architecture
Four flows. Consent on every one.
Conversation transcript
Immediate- Trigger:
- User taps 'Email me this conversation'
- Contains:
- Full transcript, cited sources, recommended resources, disclaimer.
- Consent:
- Explicit email entry + checkbox
Resource follow-up
+1 day- Trigger:
- User downloads or saves a recommended guide
- Contains:
- The guide, one related lesson, and a soft path to a strategy session.
- Consent:
- Marketing opt-in required
Escalation handoff
Immediate to user, instant alert to Tyler- Trigger:
- Conversation escalated to Tyler
- Contains:
- What was asked, what AI Tyler could not answer, suggested prep.
- Consent:
- Implied by escalation request
Education track
Weekly, max 6 emails- Trigger:
- Three or more conversations in a journey stage
- Contains:
- Stage-matched lessons from the BUFF Academy. One-click unsubscribe.
- Consent:
- Marketing opt-in required
Implementation sequence
What gets built, in what order.
The experience is designed first so the model is the last variable, not the first.
- Phase A · Now
Experience design
Conversation UI, citation and confidence contracts, escalation, admin surface. Scripted responses.
- Phase B
Knowledge base
Cloud tables for documents, chunks, reviews, and threads. Admin upload + human approval queue.
- Phase C
Retrieval + model
Embeddings, hybrid retrieval, layered system prompt, streaming responses, grounded citations.
- Phase D
Lifecycle
Transcript email, escalation alerts, nurture tracks, analytics on gaps and escalation reasons.
Internal
Knowledge management, upload, and review queue for the BUFF team.