UnitPulse · 2026 · Product Design Lead
Rebuilding UnitPulse's renter experience around conversational search and live listings, with property content designed for discovery through search and AI. Shipped in about six weeks.
UnitPulse is a venture incubated by Tripalink leadership.
An assistant that turns a vague need — "walkable to USC, dog-friendly, under $2,500" — into real listings on the map. Grounded on live inventory, so it never invents a home.
Inline stepped cards capture must-haves.
A live map with price pins beside the result grid.
Shortlist homes and plan several tours in one pass.
Adapts to past searches and stated context.
Engineering owned the server-rendered scaffold and data layer; I built the renter surface on top of it — and never claim it.
A renter’s search no longer starts in Google — it starts in a chat, often answered before anyone clicks. The old site couldn’t be found in either place: a client-only SPA, invisible to search engines and answer engines alike. The fix was a harder design problem than a redesign.
A client-only SPA — nothing server-rendered for Google or an answer engine to read.
Search now begins in answer engines, not only Google — and a site that can't be cited there is invisible to a growing slice of demand.
Keep the experience rich and app-like, yet fully indexable and citable — without the ephemeral chat polluting the index.
Before designing the search, I mapped how people start looking for a home. 3 starting points shaped the design, and they want different things.
Relocating on a timeline; knows the essentials.
NeedsFilters and a map — fast, precise, no conversation.
First place in a new city; knows the vibe, not the specifics.
NeedsConversation that turns "walkable and lively" into real listings.
Down to a few homes, comparing the details.
NeedsAnswers about one place — commute, schools, walkability.
By the time I shipped, conversational search had become table stakes — Zillow, Apartments.com, and Redfin all launched it in 2025–26. So I evaluated each on the three axes this rebuild competes on.
Natural-language search, an "AI mode" beta, and a ChatGPT integration — the deepest AI stack. But filters and chat stay separate modes, and pages aren't built for organic citation.
Voice + chat "Smart Search" (sitewide Jan 2026) drives big engagement lifts — but it's a guided refine layered on the portal, not unified with the filters.
Sierra-built AI Search replaces filters with pure chat — a linear Q&A. Fluid, but you lose the at-a-glance control precise renters rely on.
The key finding: AI chat is everywhere now, but it’s mostly bolted on as a chat-only Q&A — renters lose the fast, structured control they rely on. No one fuses conversation and filters on one shared state, and no one structures their pages to be cited by answer engines. That’s the opening this rebuild is built for.
Two kinds of renters, one missing interface: nobody lets you start vague, get precise, and keep asking — all on the same canvas.
So I designed the two models as one. Conversational handles intent; faceted handles precision; both write to a single shared state, so they can never drift apart or contradict each other.
Renters don’t search in fields. They say what they want, change their mind halfway, and ask about the home in front of them. The copilot has to keep up with all of that and still land on a real next step.
| Aspect | Filters alone | A conversation you can steer |
|---|---|---|
| Context-aware | Every change starts over from the form. | It remembers your filters and the home you're viewing. “Widen by $200” edits the saved budget. |
| Speaks your language | You translate what you need into fields. | One message fills several: “2b2b under $3,000, move in Aug 23.” |
| Relevant to you | Every listing, every field, every time. | Questions about one home are answered from that home's own data, in a few sentences. |
| Honest when nothing fits | “0 results.” | It says so in your terms and offers up to 3 one-tap ways to widen. It never loosens a filter on its own. |
| Built for action | A contact form at the bottom of the page. | Ask to tour and the form opens in the chat. It never says a tour is booked. |
Context
“Can you go $200 higher on my budget?”
Listen between the lines
“I also have a dog.”
Move the conversation forward
“2 bedrooms under $1,200 in Los Angeles.”
Turn-taking
“Can I see this one on Saturday?”
Guidance vs control Revisit next
“What can you tell me about this building?”
Move the conversation forward
Renter: 2 bedrooms under $1,200 in Los Angeles
Great news, here's what I found for you!
System: 0 homes
Context
Renter: Widen my budget by $200
System: The model works out the new budget
Updated your budget.
Listen between the lines
Renter: Tell me about The Linden
Searching for homes near Linden Ave…
Move the conversation forward
Assistant: Here are 3 homes.
Renter: Any bedrooms
Here are 3 homes.
Assistant: I couldn't find any 2-bedroom rentals in Los Angeles under $1,200/mo right now.
System: Each chip changes the filters in code
Control One rough edge left
Renter: Remove bedroom filter
Here are 2-bedroom homes in Los Angeles.
Listen between the lines Open question
Renter: Can I keep a cat?
Yes, cats are welcome here.
Designing for the answer engine — not just the search engine — is a bet most rental sites haven’t made.
Every home gets one stable page that Google can index and an answer engine can quote.
Each property page carries real questions — “Is it walkable to USC?” — exactly what a renter asks an engine, structured so it can lift them.
Chats stay private and out of search results; old links redirect to each home's page.
"How close is it to USC?"
About a mile from campus, with two bus lines at the corner.
Apartments near USC — studios
unitpulse.ai › ca › los-angeles › …
About a mile from campus, two bus lines…
"…about a mile from USC, with two bus lines nearby."
unitpulse.aiTo be precise: this is architecture and intent designed for citation, not a measured count of AI citations. Classic indexing is the part I can measure today.
The redesign was a prioritization decision, not a coat of paint — knowing what to leave out so the AI bet could ship.
A named brand-token palette, a serif display + grotesque body, a bento home — so the AI features read as premium, not experimental.
44pt touch targets on mobile, 36–40px on desktop, aria-labels on icon buttons — defaults, not retrofits.
01Discover
02Define
03Develop
04Deliver
An inquire modal — name plus email-or-phone, either is fine.
A date and time-slot grid, with SMS consent.
Twilio A2P 10DLC consent + disclosure; GA4 fires only after a real 2xx — so the count tracks real leads, not optimistic clicks.
A historical suite in English and Chinese, each case judged on how the message was routed, not just the visible reply.
A real signal, not a green-washed “all passed.” Mapped back to 7 known prior bugs as a regression guard.
Adversarial cases proved the router must refuse protected-class filters, never extract them, and redirect to a lawful search.
47 cases · 9 intent families · English + Chinese · each run in a fresh session and judged on how the message was routed, not just the visible reply.
The suite caught a real Blocker: one path extracted a protected-class attribute as a filter and affirmed it; another path handled it as the gold standard. Compliance must be tested, not assumed.
Checked before anything else. The copilot declines in 2 or 3 plain sentences, offers lawful criteria and clears every saved filter.
“Koreatown” is a place and “Is this area safe for families?” is a safety question. Neither is flagged, and area answers carry no crime or demographic data.
A 6-week-old production rebuild, live and ramping.
These are early numbers for a 6-week rebuild. What matters is that the architecture is live, indexing is climbing from zero, and the funnel is converting real leads. Applications aren’t flowing yet (0 in the window), and I’d rather say that than dress it up.
Inquiries and tour requests
Today96 in the first 6 weeks: 54 inquiries, 42 tour requests
I'd hold it toTrack it as a rate per chat. If it drops 20% below its 4-week average, check the latest flow change first.
Messages to reach a request
TodayNot measured yet
I'd hold it toIf renters need more than about 8 messages to send one, revisit the intake form and the no-results replies.
Every release has to pass
The forward move was to build content an AI can cite, and to prove an AI-native rebuild can ship in weeks, in production, without sacrificing craft, an honest index, or compliance.
The testing is the part I’d carry anywhere: define the intent space, automate the evaluation, read the logs, and treat Fair-Housing compliance as a first-class thing you verify, not assume. It also taught me to test the handoff between the interface and the AI flow. The answer can sound reasonable while the product has lost the user’s context.
AI helps me try ideas and build faster. I still need to define the behavior, review the details, and test whether it works for the person using it.
Here, asking “Does it allow dogs?” about the home you’re viewing adds a Pet friendly filter. In our leasing chat, asking adds nothing; only “I need parking” does. I’d pick one rule and test it with renters.
