2023 — 2024
Voice AI platform
Calls from a contact database, scripted dialogue with GPT and lead qualification. Qualified leads passed to sales, stage-by-stage drop-off analytics and A/B testing of scripts.
Lead full-stack / AI developer
The challenge
Complex dialogues needed to be managed, and scenarios compared using the results of actual conversations.
What I built
- Developed the voice loop with speech recognition, GPT and speech synthesis.
- Built a Laravel/Vue platform for contacts, campaigns and calls.
- Designed a versioned dialogue graph editor.
- Built a funnel showing qualified and dropped contacts at each dialogue stage; compared scripts and voices through A/B testing.
- Connected the contact database, calls, qualification and handoff to a sales call; ended conversations politely for unqualified contacts.
Results
- Built a platform with a voice agent, scenario editor, CRM features and analytics.
- Handed the system over to the next developer.
Components & connections
Voice Input
- Phone conversationAudio input and spoken responses.
- Speech-to-textTranscribe spoken input.
- Text-to-speechRender spoken responses.
- End · conversationPolitely end the conversation with an unqualified contact.
Voice Ai
- Versioned scenario graphConversation states and transitions.
- GPT · DialogueInterpretation and response generation within the configured script.
- Lead · qualificationEvaluate whether a contact matches the script criteria.
Voice Product
- Admin & CRMContacts, scenarios and campaign management.
- Contact databaseContacts and campaigns for outbound calls.
- Funnel · Conversions and A/BQualified and dropped contacts by stage, with script and voice comparisons.
- Sales call · Sales handoffHandoff of a qualified contact to the sales workflow.
- Phone conversation→Speech-to-textaudio
- Speech-to-text→Versioned scenario graphutterance
- Versioned scenario graph↔GPT · Dialoguecontext
- GPT · Dialogue→Text-to-speechreply
- Text-to-speech→Phone conversationvoice
- Admin & CRM↔Versioned scenario graphversions
- Versioned scenario graph→Funnel · Conversions and A/Bevents
- Contact database→Phone conversationoutbound calls
- Versioned scenario graph→Lead · qualificationscript criteria
- Lead · qualification→Sales call · Sales handoffqualified
- Lead · qualification→End · conversationunqualified
- Lead · qualification→Funnel · Conversions and A/Bstage and drop-off
- Funnel · Conversions and A/B→Versioned scenario graphA/B script selection
Engineering takeaway
Context
A voice automation product for the US market.
- Conversations are sensitive to latency.
- Dependence on external speech and AI services.
Approach
Connected the voice agent to a dialogue graph and feedback from analytics.
- A transition graph instead of a linear rule list.
- Separate call execution, scenario editing and analytics.
- Compare scripts using conversions and drop-off at each dialogue stage.
Lessons
- Scenario versions and analytics make agent behaviour easier to evolve in a controlled way.
Stack
- Python
- GPT API
- Cloud STT/TTS
- Laravel
- Vue
- Graph editor
- CRM