AI Architect (Voice AI)

• Real-time voice pipelines: streaming STT, turn handling, low-latency LLM inference - hands-on

• LLM engineering: prompt engineering, structured extraction, guardrails, model A/B evaluation

• Observability and evals: Langfuse or similar; golden datasets; latency, accuracy, and cost dashboards

• AWS: Bedrock, serverless patterns, SES; token economics and per-call cost engineering

• Full-stack pragmatism: strong Python; enough TypeScript / Chrome-extension knowledge to own the integration

• Clear spoken and written English for demanding US executives

• Contact-center / agent-assist patterns and metrics (handle time, cost per call, concurrency)

• Production LLM operations: load testing, data isolation, incident handling

Key characteristics (screen for all four):

1. Voice AI in production - mandatory. Shipped at least one real-time voice or speech product to real users (agent assist, voice bot, live transcription copilot). Candidates will demo real artifacts at the interview.

2. 6+ years hands-on AI/ML engineering, with strong recent LLM production practice

3. Latency and reliability record. Can show measured P95 reductions and concurrency fixes on a live system

4. Consulting / client-facing seniority. Calm and precise under detailed UAT scrutiny; manages expectations well

• Nice to have: empathy-sensitive domains (healthcare, veterinary, insurance) and PE-sponsored rollouts

Nice to have:

• Chrome extension delivery; telephony / streaming stacks (Amazon Connect, Twilio, LiveKit)

• Langfuse in production

• US client experience with Eastern-time overlap

The client is the largest US network of in-home veterinary hospice and end-of-life care. A major US private-equity sponsor drives the AI program and plans more projects across its portfolio.

We built a real-time voice copilot for their Veterinary Care Coordinators (VCCs). The copilot listens to live calls with pet families. It extracts appointment and clinical fields while the call runs. It fills the client's scheduling system through a Chrome extension. A second workstream, the Vet Visit Copilot, sends each vet an AI pre-visit briefing by email (Amazon SES).

Next is the production phase.

Stage: production SOW in executive alignment; start expected September 2026.

Duration: multi-month, with strong extension probability. 0.5 FTE minimum; ramp toward 1.0 FTE as production scales.

Why the role is open: the current architect moves to another strategic build. He stays at 0.15-0.2 FTE for supervision and knowledge transfer during ramp-up, so the new architect gets a structured handover.

• Own the technical architecture and delivery of the voice copilot from validated PoC to production

• Hit the bar this client tests against: latency, accuracy, concurrency, and cost

• Keep expectations aligned: production polish is in scope now; protect the team from silent scope creep

• Transfer knowledge continuously to the client's team and Neurons Lab engineers

• Own the full pipeline: streaming speech-to-text, LLM field extraction, Chrome-extension delivery, and AWS infrastructure

• Drive latency work: cut P95 from ~6s toward ~2s; remove post-processing corner cases (occasional ~1min lag on one field type)

• Run model A/B tests (current pair: Claude Haiku vs GPT Luna) with golden-set evaluation for phonetic name and email accuracy

• Own evaluation and cost: Langfuse traces, accuracy dashboards, real per-call cost from live calls, and an optimization plan

• Harden for production: 5-10+ concurrent calls, strict data isolation between users, monitoring, alerting, and safe rollback

• Ship epics end to end (example: the SES email briefing service); always keep a demo fallback so a live session never fails

• Front technical discussions with a meticulous client; VCCs test edge cases and expect production quality

• Present concrete system behavior, with numbers - this account rewards evidence, not slides

• Hold the scope line: tie every feedback item to the SOW; route roadmap items (learning loop, persistent memory) to future phases

• Keep internal discussions internal; all client-facing materials pass ADM review before sending