Technical Advisory Brief — AI-Powered Call Center Optimization Platform

# Technical Advisory Brief — AI-Powered Call Center Optimization Platform **Engagement type:** Advisory / consulting (not implementation — we have an in-house dev workflow) **Location:** Remote. Client based in Santiago, Chile (GMT-4) **Languages:** English or Spanish **Confidentiality:** NDA required before access to code, client names, or commercial details --- ## 1. Project Overview We are building a call center optimization platform for the pharmacy benefits division of a large Chilean retail chain (name disclosed under NDA). The platform replaces a traditional 1-to-1 voice call model with a 1-to-many model: - Field staff at retail branches call an IVR line - A voice bot captures structured case data (IDs via DTMF, free-text reason via speech) - An AI layer synthesizes each case into a triage card (condition, medications, urgency, suggested action) - Back-office agents handle up to 4 simultaneous text-based chats from a multichat console, with AI-generated context per case A working demo exists and a pilot with the end client is approved. We are now moving from browser-based demo architecture to production telephony, and we need specialist advisors to pressure-test our decisions before we commit budget. ## 2. Current Stack | Layer | Technology | |---|---| | Frontend | React 18, TypeScript, Vite, Tailwind, shadcn/ui, Zustand | | Backend | Supabase (Postgres, Realtime subscriptions, Edge Functions) | | TTS | ElevenLabs via Edge Function (production-grade voice, es-CL) | | STT (demo) | Web Speech API (browser) — must migrate to server-side for production | | AI triage | LLM via Edge Function (async synthesis, out of the real-time loop by design) | | Telephony (existing client infra) | Asterisk + commercial dialer over private VPN | | Telephony (production target) | Twilio | **Architecture decisions already made (not up for debate, but you may flag risks):** - Hold-based turn model — no live bidirectional voice mediation - Escalation handled via direct outbound call, no SIP transfer - LLM runs asynchronously during hold, never in the real-time audio path - No patient database in production; national ID passes as plain text ## 3. Advisory Modules (apply to one or more) ### Module A — Twilio Telephony Architecture We need an advisor with hands-on production experience in: - Twilio Programmable Voice: IVR design, DTMF capture via the Gather verb, call forwarding to DID with metadata passthrough - Integrating Twilio with an existing Asterisk-based IVR (bypass DID for pilot, native forward option for production) - Twilio Media Streams for real-time audio capture feeding server-side STT - Latency budgeting and cost modeling for a ~contact-center-scale deployment in Chile **Deliverables:** written review of our 10-week integration plan, identification of failure points, recommended call flow diagrams, 2–3 video sessions. ### Module B — Server-Side STT for Chilean Spanish (Medical Domain) - Comparative recommendation: Google STT / Azure / Deepgram / Whisper-based options for es-CL with custom vocabulary (medication names, clinical terms) - Strategy for custom vocabulary / phrase hints / post-processing (we already have a normalization pipeline: diacritics → acronym whitelist → collision blacklist → Levenshtein) - Accuracy benchmarking methodology we can run against real recorded calls - Streaming vs. batch trade-offs given our hold-based turn model **Deliverables:** written vendor recommendation with cost projections, benchmark protocol, 1–2 video sessions. ### Module C — Conversational AI Platform Evaluation - Experience deploying ElevenLabs Conversational AI (or comparable: Vapi, Retell, Bland) in production telephony - Specifically: whether the platform supports hold/pause/resume turn architecture, and integration patterns with Twilio - Cost-at-scale analysis: managed conversational AI vs. custom bot (we have internal estimates and want them challenged) **Deliverables:** written feasibility assessment, architecture recommendation for a phased migration (managed platform → custom bot), 1–2 video sessions. ### Module D — Supabase / Edge Functions Production Readiness - Review of Realtime subscription architecture under concurrent multichat load - Edge Function cold start mitigation (we observe ~2.6s first invocation on TTS) - Security review: RLS policies, secrets handling, audit logging for a regulated-data context **Deliverables:** written architecture review with prioritized findings, 1 video session. ## 4. Engagement Terms - **Format:** blocks of 5–10 hours per module; written deliverables required (not calls only) - **Trial:** first engagement is a paid 2-hour review of one document we provide; continuation depends on quality - **No repo access initially.** Advisors work from architecture docs and diagrams. Code access only after NDA and demonstrated fit - **IP:** all deliverables are work-for-hire; advisor may not reference the client or project publicly ## 5. What We're Looking For (screening criteria) - Demonstrable production experience — name specific projects, call volumes, and platforms (we will ask follow-up questions) - For Module A: at least one production Twilio Voice deployment with IVR + forwarding; Media Streams experience is a strong plus - For Module B: experience with Spanish-language STT tuning; LATAM accents a strong plus - Comfort giving direct, critical feedback. We want our plan attacked, not validated - Availability for video calls overlapping GMT-4 business hours ## 6. How to Apply Reply with: 1. Which module(s) you're applying for 2. One paragraph describing the most similar production system you've worked on (platform, scale, your role) 3. One specific risk you'd flag in the architecture described above (this is the real filter) 4. Your hourly rate and availability