Voice AI, Prompts, and Memory for a Real Estate Platform

Industry

Real Estate Technology, Software as a Service

Geography

United States

Cooperation Period

4 Months

Summary

The client operates an AI platform for real estate teams, combining lead engagement, agent routing, and agent support into a single stack that installs on any CRM, dialer, and calendar the team already uses. Clixlogix engaged as an AI partner covering both advisory and delivery work. Advisory covered AI architecture and tech stack recommendations across the platform’s core capabilities. Delivery covered 3 AI workstreams (voice conversation pipeline, prompt library, and memory architecture) and 1 integration workstream (connector fabric across 8 real estate CRMs, 3 dialers, and 3 calendar ecosystems). The platform now runs across more than 800 real estate teams and 7,000 agents in the United States.

The Client

The client operates in the United States and runs an AI platform for real estate teams. Its founders come from senior product and engineering roles at scale, with deep backgrounds in real estate technology. The platform combines AI lead engagement, agent routing, and agent support into a single stack that installs on the CRMs, dialers, and calendars real estate teams already use.

Challenge

The client’s product bet is direct. If the AI cannot hold a conversation the way a good agent would, meaning fast, contextual, warm, and appropriate for the moment in the lead’s journey, the platform does not have a business. Voice quality, prompt quality, and memory quality decide whether every AI capability the platform ships earns the trust of the team and the lead.

The client already had a strong engineering team. What the team wanted was a specialist AI partner that could

advise on the voice conversation architecture before key decisions became difficult to reverse
evaluate LLM providers against real workloads
establish a disciplined prompt system across the product
design a memory architecture capable of carrying lead context across a home search that may last for months

The engagement also carried an integration requirement. The platform had to install on any real estate CRM, any real estate dialer, and any calendar the team already used, with a single API key and no data migration. That constraint pushed the AI work and the integration work onto parallel tracks that had to stay in sync.

Requirements

The delivered work covered 4 functional areas, each with its own acceptance conditions.

Voice conversation pipeline

Sub second turn latency mouth to ear
Interruption handling that respects lead speech in progress
An interface abstract enough to let the client swap speech to text or text to speech providers without touching prompt logic
Voice activity detection tuned for the pause behavior of real estate calls

Prompt library

Versioned prompts under source control
Golden test cases per prompt with automated evaluation on every deploy
Retrieval augmented prompts drawing from named retrievers with documented behavior
Compliance guardrails composed as shared fragments across every capability

Memory architecture

Lead context retrieval spanning arbitrary time gaps up to a full home search of many months
3 parallel retrieval tracks (recency, semantic, structured) with provenance carried at every event
Documented decay policy, auditable and adjustable per tenant

Connector fabric

Single API key install with no data migration and no admin work on the CRM side
Bidirectional sync across 8 real estate CRMs, 3 dialer providers, and 3 calendar ecosystems
Common object model that abstracts each CRM’s native schema
Framework that keeps adding a new CRM a scoped implementation of the connector interface

Solution

Clixlogix engaged the client as an AI partner covering both advisory and delivery work. Advisory covered AI architecture across the platform’s AI stack and a tech stack recommendation documented against cost, latency, and quality tradeoffs. Delivery covered 3 AI workstreams (voice conversation pipeline, prompt library, and memory architecture) and 1 integration workstream (connector fabric across real estate CRMs, dialers, and calendars). Advisory ran alongside delivery for most of the engagement, and the client’s engineering team owned the platform’s core build.

Fig 1 – AI stack architecture

AI architecture consulting for the AI stack

The client’s platform runs several AI capabilities on the same technical foundation. Real time voice and text conversations with leads. Signal based prioritization and routing to the right agent. An assistant that supports agents by call and text between showings across CRM, follow up, and calendar work. Each capability carries a different latency budget, a different context window, a different tolerance for creative output, and a different failure mode when the AI gets it wrong.

Clixlogix worked alongside the client’s engineering team from an early architecture design phase to lock in how the platform’s AI capabilities would share a common foundation without collapsing into a single monolithic prompt. The recommended architecture separated 2 tiers of components.

Shared across the stack

Memory retrieval
Embedding pipeline
Evaluation harness

Owned by each capability

Voice orchestration for lead conversations
Decision logic for routing
Task orchestration for the assistant

That separation gave each capability room to evolve without disturbing the others, and it kept the shared components under one team’s ownership on the client’s side.

Advisory sessions ran on a regular cadence for the full engagement. Design reviews covered new features before the client’s team committed to a build, and postmortems covered production issues after the client’s team resolved them. That posture kept AI architecture decisions grounded in real workload behavior and let both teams learn from each other’s calls.

Model and tech stack recommendation

Provider capabilities, pricing, latency profiles, and model performance change frequently. The client needed a defensible way to decide which provider belonged behind each AI capability, based on the workloads the platform runs and the constraints the business carries.

Clixlogix ran the evaluation. Recommendations covered

LLM providers for text generation across the platform’s AI capabilities
Speech to text and text to speech providers for the phone call experience
An embedding model and a vector store for memory retrieval
An orchestration framework for LLM calls
A hosting posture for the AI components the platform depends on

Each recommendation carried tradeoff analysis on cost per conversation, latency, output quality against real prompts, vendor stability, and switching cost if the market moved. The client’s team took the recommendations into their planning cycle. The stack included GPT-4o and Claude 3.5 Sonnet for LLM routing, Deepgram Nova and ElevenLabs Turbo for the voice pipeline, text-embedding-3-large with Pinecone for memory retrieval, and LiveKit Agents for voice orchestration.

Clixlogix documented the reasoning behind every recommendation so the tradeoffs stay visible if the market moves and any decision needs revisiting. That paper trail helps the client explain AI infrastructure choices to its board and to prospective enterprise buyers with confidence.

Voice conversation pipeline

The platform’s voice conversations are the sharpest technical test it faces. A lead who calls in expects a human sounding voice that responds in the rhythm of a real conversation. Latency has to stay inside the window a human tolerates, and interruptions from the lead have to interrupt the AI mid sentence, not queue up until the AI finishes its turn.

Clixlogix designed and built the voice pipeline behind the platform’s lead conversations. The pipeline runs Deepgram Nova on the inbound audio stream, feeds partial transcripts into GPT-4o before the lead finishes speaking, streams the model’s response into ElevenLabs Turbo that begins voicing before the model has generated the full sentence, and monitors the audio stream for interruption signals that trigger an immediate cutoff and re prompt.

Fig 2 – Voice pipeline latency

Turn taking received particular attention. Early prototypes waited for a fixed timeout after the lead stopped speaking, which created premature responses and unnatural interruptions. The final pipeline uses voice activity detection tuned against real estate call behavior, a rolling estimate of the lead’s speaking rhythm, and interruption logic calibrated against the latency profile of the selected text to speech provider.

Consulting Insight

Turn detection is a model problem

The first voice pipeline prototype used a fixed silence timeout to decide when the lead had finished speaking. Real estate calls proved that logic wrong within days of production traffic. Leads paused mid sentence to check a note, mid thought while formulating a question, and mid pricing statement while pulling up a listing on a laptop. The pipeline treated each pause as a turn boundary. The AI cut them off. They stopped calling back.

Clixlogix and the client's engineering team replaced silence detection with a custom turn detection model that reads the lead's speech content, cadence, and prosody. The pipeline held its latency budget and stopped interrupting.

Transferable principle Silence duration is a proxy that fails on the calls that matter most, which are the ones where the lead is thinking hard about a real decision. A voice AI product that uses silence timeouts trades conversation quality for engineering simplicity. That trade cannot be made on a product where conversation quality is the product.

Fig 3 – Turn detection comparison

The pipeline exposes the same interface to the LLM regardless of which speech to text or text to speech provider sits behind it. That interface let the client swap text to speech providers once during the engagement without touching the platform’s voice prompts, and it lets the client add a second voice provider for regional accents when a customer segment demands it.

Prompt library across every AI capability

The platform’s AI stack drives dozens of specific behaviors across the lead lifecycle. LLM prompts classify inbound intent, draft opening responses, hold multi turn conversation, extract qualifying data, score and route leads, write agent briefings, handle calendar work, draft follow ups, generate CRM updates from voice, and answer agent market queries. Every one of these behaviors is a prompt, or a chain of prompts, that needs to be reliable, testable, and versioned.

Clixlogix built the prompt library that every AI capability draws from. Prompts live in a versioned repository the client’s team edits, deploys, and rolls back. Every prompt carries a schema for its inputs and its expected output, a set of golden test cases the prompt has to pass before it deploys, and a link to the evaluation record from its last production run. Sections of prompts that draw on retrieval, including memory recall, listing context, and agent voice profiles, pull from named retrievers with documented behavior, so a prompt author knows what will land in the context window at runtime.

The library also enforces guardrails common to real estate. Fair Housing Act constraints, a ban on financial advice, and a ban on specific promises about a property or the market appear as shared prompt fragments composed into every prompt in the library. That composition means a change to a compliance constraint updates every AI capability at once, and it keeps compliance auditable in a single place.

Consulting Insight

Compliance is horizontal

The first version of the prompt library treated compliance guardrails as prompt content. Every prompt that touched a lead conversation carried its own Fair Housing Act language, its own ban on financial advice, and its own ban on promises about a property or the market. When a compliance officer at the client asked for a small tightening of one constraint, the team had to find and edit that constraint in more than 20 prompts across the library. The team caught most of them. It missed 2. Those 2 shipped a slightly outdated compliance line to production for a week before the evaluation harness caught the drift.

Clixlogix restructured the prompt library to compose compliance from shared prompt fragments. A single fragment for Fair Housing Act constraints. A single fragment for the ban on financial advice. A single fragment for the ban on specific promises about a property or the market. Every prompt in the library composed these fragments at build time, and any change to a fragment updated every prompt that used it on the next deploy. The compliance officer got a single place to review. The library got a single place to audit.

Transferable principle Compliance in AI outputs is horizontal work. Every capability the AI ships touches it. Treating it as vertical prompt content means the library carries dozens of copies of the same constraint. The first constraint change proves it does not scale. Building shared prompt fragments for compliance at the start costs one day. Retrofitting it later costs a month.

Fig 4 – Compliance fragments before and after

Conversation memory and lead context architecture

Real estate leads travel long paths. A buyer might visit the platform in one month, drop off, come back the next, tour 3 properties later, and close 9 months in. Every touch across that span should feel connected to the lead. Every agent handoff should carry the full history. Every AI response should acknowledge what the platform already knows about this specific person.

Clixlogix designed the memory architecture that carries lead context across arbitrary time gaps. Every interaction, whether a text, a call, an email reference, a CRM update, or a showing, writes to Pinecone as an embedding from text-embedding-3-large, and to PostgreSQL as a structured event with provenance. Memory retrieval happens on 3 tracks.

Recency window pulls the most recent handful of interactions verbatim from PostgreSQL
Semantic retrieval pulls the most relevant historical interactions from Pinecone by embedding similarity to the current context
Structured lookup pulls stable facts about the lead, including name, budget, timeline, preferences, and current agent, from a canonical schema in PostgreSQL

Fig 5 – 3 track memory retrieval

Consulting Insight

Memory needs structured tracks alongside semantic retrieval

The first version of the memory architecture ran purely on embedding based semantic retrieval. The architecture indexed every past interaction, and every current conversation queried the index for the most similar prior context. In evaluation, the retrieval quality looked strong. In production, the AI started misfiring in ways evaluation had not caught. It confused 2 leads with similar first names. It cited old budget numbers a lead had explicitly revised. It surfaced context from a different agent's conversation with the same lead months earlier.

Clixlogix added recency and structured lookup tracks alongside semantic retrieval, and made sure structured facts arrived first. Every retrieved event carried provenance the retrieval scorer could weight against.

Transferable principle Semantic retrieval is powerful for surfacing thematic context. It is unreliable for the facts a real business decision runs on. Memory architectures for high stakes AI need at least one track that answers the question with a database join. Similarity scores alone will surface facts that were true 6 months ago and are false today.

Memory writes carry provenance. Each event records the source (which capability produced it, which channel it came from, which prompt version generated it), a timestamp, and a confidence signal where relevant. A fact the AI extracted from a call carries different weight than a field the agent typed into the CRM. That provenance lets memory retrieval strategies weight different sources appropriately, and it lets the client’s team audit any decision the AI made against the memory record behind it.

The memory architecture also supports pruning. Not every event stays retrievable forever, and the store applies a documented decay policy to keep retrieval quality high as leads accumulate interaction history. The decay policy is auditable, adjustable per tenant, and consistent across every AI capability that consumes memory.

Connector fabric across CRMs, dialers, and calendars

The AI capabilities mean nothing if the platform cannot install on the team’s actual stack. Real estate teams already run on CRMs like Follow Up Boss, Sierra Interactive, MoxiWorks, kvCORE, Lofty, BoldTrail, HubSpot, and Salesforce, dialers like Mojo, PhoneBurner, and Vulcan7, and calendars across Google Calendar, Microsoft Outlook, and Apple Calendar. Any friction at install kills the deal.

Clixlogix engineered the connector fabric that delivers the platform’s API key promise. A team activates the platform on its existing stack from a single API key entry, with no data migration and no admin work on the CRM side. Behind that single field, the fabric holds

A common object model across every CRM connector
A credential store that handles rotation and revocation
A sync engine that runs in both directions with conflict resolution
A webhook plus polling infrastructure that keeps lead events flowing within the AI stack’s latency budget
Dialer connectors that report call events into the platform and initiate outbound dials
Calendar connectors that read and write across the major calendar ecosystems

The fabric ships as a framework the client’s team extends. Adding a new CRM to the library becomes a scoped implementation of the connector interface. The base handles credential management, sync scheduling, webhook validation, conflict resolution, and observability. The marginal cost of the next CRM stays predictable enough for the client’s sales team to commit to a timeline during the demo.

Fig 6 – Connector fabric

Results

The AI stack Clixlogix helped design and deliver produced measurable outcomes across engagement volume, conversation quality, handoff readiness, and closed transactions. The 7 result cards below present each number with the operational context behind it.

12,800 Plus Leads Engaged Across the AI Stack

The AI stack Clixlogix delivered engaged more than 12,800 real estate leads in production conversation. Every one of those leads received a personalized response from the voice or text engagement flow, moved through the follow up cadence the memory architecture drove, and passed under the evaluation harness the prompt library ships with. That production volume gave the client's engineering team a reliable trace against which to measure future prompt changes and model swaps.

1 in 4 Engaged Leads Reach Meaningful Conversation

Of every 4 leads the AI stack engaged, 1 progressed beyond a single reply into real multi turn dialogue. Meaningful conversation counts a lead that responded more than once and stayed on topic long enough for the AI to extract qualifying data. That signal matters commercially. It marks the point at which the engagement flow has earned enough trust from the lead to move the relationship forward toward a live handoff.

1 in 3 Meaningful Conversations Advance to a Live Call

Of the meaningful conversations the AI stack held with leads, 1 in 3 escalated to a live phone call with a real estate agent. The handoff decision runs on a classifier the client's engineering team and Clixlogix tuned together, and every escalation carries a full briefing package the receiving agent sees before the call connects. That progression rate turns the top of the funnel into a stream of leads warm enough to close, without the operational overhead of a sales development team calling every inbound lead.

800 Plus Real Estate Teams in the US

The platform reached more than 800 real estate teams across the United States on the AI stack Clixlogix helped design and deliver. That footprint spans large brokerages, small teams, and solo brands, each running on its own CRM, dialer, and calendar. Every team runs the same voice pipeline, the same prompt library, and the same memory architecture behind the AI, at its own operational scale.

7,000 Plus Agents Working with the AI

The AI stack now supports more than 7,000 real estate agents in daily practice. Each of those agents interacts with the platform by voice and text between showings, receives live handoffs on warm leads, and uses the assistant for CRM updates, follow up drafting, and calendar work. Agent adoption at that scale is the acid test for any AI product built for a professional user base. Daily use is the only real measure of it.

120 Plus Live Handoffs at a Reference Team

At one reference team on the platform, the AI stack transferred more than 120 live phone calls from AI conversation to a real estate agent inside a single deployment window. Every one of those transfers carried a briefing package that summarized what the AI had learned about the lead in prior touchpoints. The agents received a warm caller who wanted to talk. The operational cost of that pipeline sat below what a comparable sales development headcount would have cost.

15 Plus Closings in Under 5 Months

Of the 120 plus live handoffs at that reference team, more than 15 leads converted into closed real estate transactions inside the first 5 months of the deployment. That closing count reflects the full contribution of the AI stack from initial lead engagement through handoff readiness. It also measures the revenue the client's commercial team tracks under a pricing model tied to closed outcomes.

Technologies and Tools

Category	Tools and Platforms
Voice Pipeline	Deepgram Nova (streaming API tuned on real estate call audio); ElevenLabs Turbo (low latency text to speech); Cartesia Sonic (text to speech fallback behind the same interface); LiveKit Agents (voice orchestration); custom turn detection model (trained on production call recordings); Twilio (inbound and outbound voice, SIP trunking for enterprise buyers)
LLM and Orchestration	OpenAI GPT-4o (primary LLM for lead conversations); Anthropic Claude 3.5 Sonnet (higher stakes LLM route); OpenAI GPT-4o mini (cost tier for high volume simple tasks); custom LLM orchestration on top of provider SDKs (unified retry, provider fallback, cost telemetry per prompt); Braintrust (prompt CI and evaluation); git (version history)
Memory Architecture	OpenAI text-embedding-3-large (embedding model); Pinecone (vector store for semantic retrieval); PostgreSQL (structured memory store, canonical lead schema); custom event log on Postgres (provenance and audit trail)
Connector Fabric	Node.js with TypeScript (connector library runtime); AWS SQS (sync engine queues, isolation per tenant); Fastify (webhook receiver with signature verification per CRM); AWS Secrets Manager (credential store, access scoped per tenant); Salesforce SDK, HubSpot API client, direct REST for real estate CRMs (Follow Up Boss, Sierra Interactive, MoxiWorks, kvCORE, Lofty, BoldTrail); Mojo API, PhoneBurner API, Vulcan7 API (dialer integration); Google Calendar API, Microsoft Graph, CalDAV for Apple (calendar integration)
Cloud and Infrastructure	Amazon Web Services (cloud); ECS Fargate (hosting for stateless services); PostgreSQL on Amazon RDS (data store); Redis on ElastiCache (cache for hot memory and prompt output); Datadog (observability for metrics, logs, and traces)
Evaluation and Quality	Braintrust (prompt evaluation with golden test cases per prompt, run on every deploy); Braintrust LLM as judge configurations plus custom real estate rubrics; custom UI on Braintrust exports (human review workflow for edge cases the AI flagged low confidence on)

Services Delivered

AI Architecture Consulting, AI Software & Application Development, Voice AI Engineering, Prompt Engineering, Conversation Memory Architecture, AI Integration Services

Team Composition

AI Product Architect, LLM Engineers, Voice AI Engineers, Backend & Integration Engineers, MLOps Engineer, Consulting Advisors

Have a question for our team or need help with your project?

Our team can share client references, scope your project, and answer any question about your delivery.

Related Case Studies

More engagements where our delivery teams shipped similar outcomes for clients across industries. Read on for context on the patterns we reused, the trade offs we navigated, and the metrics that landed in production.