Voice AI engineer.
I build real-time enterprise phone agents in production.
10,000+ live calls daily.
15 years in client-facing roles.
Production systems shipped fast.
Voice AI
Production voice agents and live demos
Retell AI Voice Receptionist
Live inbound voice demo with real-time visualization
THE WORK
- •Real inbound call flow with word-based session binding
- •Live transcript + agent reasoning + tool-call stream
- •Custom function calling end-to-end
- •In-memory event bus with pre-bind webhook buffering
- •Deployed with sub-500ms webhook-to-browser latency
TECH STACK
IMPACT
AI Voice Agent
Multi-tenant voice AI platform with real-time conversations
THE WORK
- •Real-time voice AI with ~600-800ms response latency
- •WebSocket streaming: Deepgram STT → Groq LLM → Cartesia TTS
- •Auto-scaling on Fly.io (0→1 in <5 seconds)
- •AI provider auto-switching with Groq/Gemini fallback
- •Multi-tenant architecture with isolated data and webhooks
TECH STACK
IMPACT
Adjacent AI Work
Infra and foundations from voice AI adjacent builds
RAG Chatbot
Production RAG with vector search and GPT-4o
THE WORK
- •PostgreSQL pgvector with HNSW indexing for semantic search
- •OpenAI embeddings (text-embedding-3-small) + GPT-4o chat
- •Semantic chunking: 500 tokens with 50-token overlap
- •Digital Ocean upload service for 25MB file processing
- •Optimized for speed: 3-5 second responses, concise answers
TECH STACK
IMPACT
Anthropic Models Manager
Centralized AI model management serving 10+ production apps
THE WORK
- •Fully automated model detection via daily cron (zero-touch updates)
- •TypeScript-based API serving live model configs to production apps
- •A/B testing dashboard with usage analytics and cost monitoring
- •Vercel Edge Functions + Neon PostgreSQL architecture
- •Public demo with clean, documented codebase
TECH STACK
IMPACT
Bland Flow
Community tool for visualizing Bland.ai voice agent pathways
THE WORK
- •Interactive flowchart visualization with React Flow
- •Shareable links with Vercel Blob storage (never expire)
- •Auto-parses Bland.ai JSON exports into visual node graphs
- •Color-coded node types: Start, End, Webhook, Route, Custom Tool
- •Click nodes to inspect prompts, conditions, and variable extraction
TECH STACK
IMPACT
Personal SMS & Voice Platform
AI-powered messaging and WebRTC calling for voip.ms
THE WORK
- •AI reply drafting with Claude (learns relationship context and tone)
- •WebRTC browser calling via FreeSWITCH SIP gateway
- •Voicemail transcription with audio playback
- •Privacy mode for masking sensitive data during screen sharing
- •TypeScript Edge Functions with CSRF and brute-force protection
TECH STACK
IMPACT
Hard Lessons
Challenges that shaped how I think about development
Industry-Specific STT Corrections
One framework, any vertical
- •Phone audio at 8kHz degrades STT accuracy
- •Domain jargon mishears as garbage ("quogged" instead of "clogged" on plumbing calls)
- •Keyword boosting catches some misheard words, not all
- •Post-process every transcript through a regex correction map
- •Map grows from real call transcripts, not guessed vocabulary
- •Sub-millisecond latency — no LLM, no API call, just regex
- •Sub-millisecond correction, stable across thousands of calls
- •Framework ports cleanly to any vertical (HVAC, medical, legal, trades)
- •Verticalizing voice AI becomes a repeatable playbook
AI-Assisted Development
Questioning TypeScript convention
- •Convention says TypeScript prevents bugs
- •Does that apply when AI writes the code?
- •Needed data, not dogma
- •Analyzed actual bugs: column renames, event bubbling, env vars
- •Pattern: none were type errors
- •All were logic/conceptual mistakes
- •Informed decision: JavaScript over TypeScript
- •Faster iteration without compile steps
- •AI bugs need investigation, not type checking
Tech Stack & Tools
Technologies I use to build, deploy, and manage production systems
Voice AI Stack
Cloud & Infrastructure
APIs & Integrations
Databases & DevOps
AI & Development Tools
The Combination
Most technical roles need both customer communication and technical execution.
Most people only have one.
I have both.
Customer-Facing (15 Years)
Understanding What Customers NeedPhone Ai Engineer | Independent (2024-Present)
Building production voice AI. Greenfield and enterprise legacy.
Co-Founder & CTO | PFM Media (2020-2024)
Built and operated digital marketing agency with real-time call center infrastructure serving real estate clients. Trained sales teams, managed client relationships end-to-end.
Behavioral Recognition | Pearson Airport (2012-2018)
High-stakes decision-making at Canada's busiest airport. Eye on Safety award for life-saving intervention.
Skills Developed:
- Explaining technical complexity to non-technical stakeholders
- Managing customer relationships through implementation challenges
- Troubleshooting under pressure with incomplete information
Technical Execution
Building What They NeedVoice AI Platforms:
Retell AI, Bland.ai, custom STT→LLM→TTS pipelines (Deepgram, Groq, Cartesia), function calling, webhook observability
Telephony & Real-Time:
Twilio Media Streams, voip.ms SIP trunking, WebRTC browser calling, sub-second end-to-end latency, half-duplex turn-taking
Production Infrastructure:
Node.js, PostgreSQL, Server-Sent Events, Caddy + PM2 on Digital Ocean, Fly.io auto-scaling, Vercel Edge Functions
What This Proves:
- Production voice AI shipped, not demos
- Pipeline debugging under sub-second latency budgets
- Platform-agnostic: Retell, Bland, or self-hosted stack
- Legacy PBX/SIP meets modern LLMs without rip-and-replace
- Real call traffic, not tutorials
- Sole engineer — I built it and I'll maintain it
This combination is rare and valuable for customer-facing technical roles:
Let's Build Something
Interested in working together? Get in touch.
Built with modern tools and AI partnership
© 2025 Cameron O'Brien