Introducing Ava
The St. Petersburg Housing Authority’s Digital Assistant
She answers from what SPHA has published, takes down a request when there is one, routes it with its context, and follows it up — and she brings in staff whenever a person is the right answer, not only when she is stuck.
The questions that fill the phone line.
The things residents call about most, in the words they use. Each one opens the assistant and puts the question to Ava.
None of these are scripted. Each answer is built against SPHA’s published pages at the moment the question is asked.
Where the answers come from.
Ava is not a search box, and she is not a person — she says so when anyone asks. She reads SPHA’s own pages and tells a resident what they say.
SPHA’s published pages
stpeteha.org, re-read whenever staff ask for it, plus whatever they add on top — policies, notices, PDFs. Nothing else is used to answer.
Waiting lists checked live
Openings change faster than any crawl. The Apply for Housing page is read at the moment the question is asked, and what it says overrides everything else.
Staff, with the context already gathered
Anything about a household’s own application, balance or account belongs with staff, and reaches them with the question and the details attached rather than as a name and a number to call back cold.
And what she refuses to do.
She cannot see an application, a rent balance or an account, and will never ask for a Social Security number, date of birth or payment details.
She will not decide eligibility, estimate a waiting time, or predict when a list will open.
She will not give legal advice or quote a figure SPHA has not published. When she does not know, she says so and gives the phone number.
SPHA staff operate it. Not a vendor.
Ava is the part residents see. Everything that decides what she says lives in a console SPHA signs into — no developer, no redeploy, no waiting on anyone to change a sentence.
A knowledge base you control
Re-crawl stpeteha.org on demand, add individual pages, upload PDFs and Word documents, or paste text outright. Weight the pages that matter most.
Words you approve
The instructions, the greeting and the suggested questions are versioned and editable. A change you regret is one click back.
Every conversation, and what it missed
Full transcripts, volume and cost by day, and a standing list of questions it could not answer — each one a gap in what SPHA has published.
Contact requests, captured
When someone needs a person, they leave their name and number without leaving the conversation. Those land in an inbox for staff to work through.
Your choice of model
Claude, OpenAI or Brooklyn. Keys are encrypted, connections are tested with a real request, and if one provider fails the next answers without the resident noticing.
Privacy built in, not bolted on
Social Security and card numbers are stripped before anything is stored. Transcripts age out on a retention window SPHA sets. Only sites you list can embed the widget.
The same Ava, out loud.
The line answers today, and the microphone on the assistant page is live. What is still being wired is the part that matters most: making the spoken answers come from the same SPHA pages the chat reads. When that lands, a resident who asks at nine in the evening gets the same answer as one who types at nine in the morning — and anything account-specific still goes to a person.
(727) 914-5168The line it answers on
One source of truth
Voice and chat answer from the same published SPHA pages. Update the knowledge base once and both channels change together.
Nobody waits on hold
The Monday morning queue, the middle of a Wednesday, a Sunday night — a caller gets an answer instead of a ring tone or a voicemail box.
Staff, already briefed
A caller who should be speaking to a person reaches staff with the question and the details already captured, rather than as a name and a number to call back cold.
Answering the question is the entry point.
NoralAI builds and operates managed AI systems — digital workers that take a piece of work from first contact through to the outcome. This assistant is one of those workers, configured for a housing authority’s front door. The same platform carries the work further.
Capture
Information pulled out of the conversation and structured, rather than left in a transcript for someone to read later.
Route
The request reaches the right team with its context attached, instead of arriving as a name and a callback number.
Follow up
Confirmations, reminders and re-contact run on their own, so nothing stalls because a person got busy.
Workflow
Tasks created, assigned and chased from what happened in the conversation.
Orchestration
Existing systems triggered in concert — no rip-and-replace of what you run now.
Managed
Designed, deployed and operated by NoralAI. There is no internal AI team to hire.
What it takes to put this in front of residents.
Everything above is running now. Three things turn it from a demonstration into a service.
One line of script on stpeteha.org
The widget is a single tag in the site template. Only domains SPHA lists are allowed to load it.
Staff sign-off on the wording
The instructions and the greeting are SPHA’s words, editable in the console. Someone at SPHA should read them before a resident does.
A retention window
How long transcripts are kept is a policy decision rather than a technical one. The console enforces whatever SPHA chooses.
Ask Ava something.
Available at any hour, in whatever words a resident has. When a person is the right answer, she brings one in with the context already gathered.