Project Overview
My Role
Lead Conversational AI Engineer responsible for developing low-latency SIP/VOIP integration pathways, structuring state-aware dialog routing chains, and optimizing semantic knowledge database lookups.
The Problem Context
Local service operations lose substantial booking opportunities due to unanswered calls during high-volume hours, administrative backlogs, or simple after-hours business closures. Standard interactive voice response (IVR) phone menus create friction and cause prospective clients to hang up.
- Static menu trees restrict customer inquiries, forcing manual staff follow-ups.
- Generic call centers cannot read or write to live WordPress calendars in real time, causing double-bookings.
- High round-trip latency in speech-to-text-to-speech pipelines disrupts natural conversational rhythms.
We built Echo to maintain low voice-to-voice latencies by routing streaming audio blocks directly through WebSockets and executing parallelized model evaluations directly linked with the WordPress database.
Key Features & Workflow
Natural Voice Interaction
Answers incoming calls, processes spoken scheduling requests, and resolves frequently asked business questions using an integrated knowledge base.
Live API Synchronization
Queries and reserves active calendar slots inside the WordPress database on the fly without requiring human operator intervention.
Missed Call Recovery
Dispatches SMS follow-up links and online scheduling alternatives instantly to mobile callers if a line remains busy.
Technical Architecture
How the system components fit together.
SIP Trunk Signal Pipeline
Establishes low-overhead, bi-directional audio streaming using WebSockets linked with standard telecom gateways.
Semantic Knowledge Routing
Parses caller intents through bounded system prompts and semantic search layers, bypassing high-cost API reasoning loops.
Atomic Database Locking
Protects concurrent calendar modifications inside WordPress, preventing double-allocations during multiple simultaneous calls.
Design Decisions
To keep the phone call experience as intuitive as possible, we focused on latency optimization, parallelizing the text-to-speech rendering pipeline while model outputs are still compiling.
- Configured chunked, sentence-level speech streaming so synthesizer audio begins playing while downstream responses generate.
- Built a localized semantic caching layer for highly repetitive questions (such as directions, hours, and rates) to minimize inference costs.
- Created fail-safe routing parameters to transparently hand off caller connections to live physical operators when threshold limits are breached.
Future Roadmap
- Dynamic payment collection processing during active telephone calls.
- Enhanced out-of-boundary call forwarding utilizing standard SIP REFER commands.
- Custom voice clone model generation from user-submitted reference WAV files.
Technology Stack
Core technologies used to build this project.
Capabilities Demonstrated
Key areas of expertise highlighted by this project.