Nathaniel Orange

Challenge

A small travel agency competes against large booking engines while running on manual quoting and follow-up. Deals need to stay current, group trips need coordination across many travelers, and booking questions arrive around the clock when no agent is available.

Role

Sole developer. Designed and implemented the application architecture, the campaign and booking-assistant subsystems, the voice layer, and the AWS and Vercel infrastructure around them.

What I built

• Group campaign system with per-campaign landing pages, briefs, email broadcast and scheduling, idea voting, and analytics events
• Booking assistant with guest flows: availability checks, callbacks, draft state, continue-later, reminders, and agent claim/dismiss handling
• Operator surfaces including an operator chat copilot, directive review and apply, and booking-change acknowledgement
• Realtime voice sessions over the OpenAI Realtime API, including a hybrid mode that uses the realtime model purely for speech-to-text and text-to-speech while reasoning runs through the standard chat pipeline
• Automated cruise deal refresh that stores a generated payload in DynamoDB so production reads never scrape or call AI providers during page render

Verifiable in the source

• Voice endpoints for session minting, hybrid STT/TTS sessions, tool dispatch, and TTS under app/api/voice
• Booking-assistant API surface spanning availability, claim, callback, reminders, drafts, and custom email composition
• Campaign API covering briefs, landing pages, email scheduling, directives, and page-view analytics
• Deal refresh pipeline documented in the repository, writing to a DynamoDB app cache table

Challenge

A pawn shop loses customers to unanswered calls and after-hours drift, and has no online inventory or appraisal path. Staff attendance and item intake are tracked manually.

Role

Sole developer. Built the voice relay server, the Next.js application and its API layer, the DynamoDB data model, and the owner and staff interfaces.

What I built

• Fastify voice relay bridging Twilio Media Streams to the OpenAI Realtime API over WebSocket, with interruption handling, audio truncation, and a mark queue to keep playback in sync
• Voice tools the model calls during a live call: schedule_visit with field validation and audit persistence, and check_store_status for real-time open/closed answers instead of guessing from the prompt
• AI appraisal flow accepting up to six labeled photos, returning condition, material and purity assessment against live metal spot prices, and creating a lead from every appraisal
• Owner dashboard with leads, staff activity, inventory management, conversation review grouped by customer and case, agent configuration, and staged data resets
• Staff clock-in with daily rotating SHA-256 QR tokens, server-side timestamps, and automated compliance flagging for invalid tokens, long shifts, and missed clock-outs

Verifiable in the source

• 674-line realtime voice server (backend/realtime_voice/server.js) implementing the Twilio media-stream bridge, tool dispatch, and interruption handling
• Twilio voice and message webhooks plus a realtime-session endpoint issuing ephemeral browser credentials
• Twenty API routes including appraise, inventory search, leads, schedule, staff-log, store-status, and agent-config
• Repository FEATURES.md maintained with a code-accurate corrections section

Challenge

A solo operator cannot answer the phone while working on a device. Calls go to voicemail, leads go cold, and there is no record of who tried to reach the business or what they wanted.

Role

Sole developer. Built the voice relay, the Next.js application, the AWS Lambda SMS pipeline, the DynamoDB schema, and the owner dashboard.

What I built

• Realtime voice relay connecting Twilio Media Streams to the OpenAI Realtime API, with interrupt handling and owner-configured voice selection
• Twilio Programmable Voice webhook path using speech gathering with OpenAI TTS playback and an Amazon Polly fallback when TTS fails
• Owner dashboard with six status modes that change agent behavior mid-conversation, configurable agent name and voice, and owner bulletins the agent weaves into replies
• Appointment booking into DynamoDB with collision checks against existing bookings on the requested date
• Parallel AWS Lambda pipeline for SMS and voice via API Gateway, sharing the same DynamoDB tables and OpenAI function schemas as the web path

Verifiable in the source

• Realtime voice relay server implementing the Twilio media-stream bridge with interrupt_response enabled
• API routes for twilio-voice, twilio-sms, chat, tts and tts/stream, state, leads, chat-logs, and realtime-session
• Lambda handlers dispatcher.py, state_manager.py, and scheduler.py with deployment and IAM setup documented
• Connection and end-to-end test scripts for OpenAI, Twilio, and DynamoDB with recorded results