AI Architect: Voice, Vision & IoT Ops Platform

We run 16 hotel and resort properties, and we are building the operations brain that runs them. We need one person who can architect the whole thing, not five people who each own a slice of it. Today our properties run on walkie-talkies, WhatsApp groups, a 9-year-old PMS on SQL Server, a Django admin panel a contractor left us, 400+ CCTV cameras nobody watches, and a maintenance log that is genuinely still a clipboard. Every one of those is a data source. None of them talk to each other. We have board approval and budget to build Aventis OS: a single real-time operations platform that ingests what is actually happening in a building (voice, video, sensors, work orders) and acts on it without a human dispatcher sitting in the middle. We have an internal backend dev and a mobile dev. We do not have an architect, and we have learned the hard way that hiring five specialists who each hand off to the next does not produce a system, it produces integration tickets. So we are hiring one senior person to own the architecture end to end, build the hard parts personally, and lead our two in-house engineers plus any contractors we add. What we are building (8 workstreams) 1. Voice: replace the walkie-talkies Staff radios and a mobile push-to-talk stream PCM audio into the platform over WebSockets. We need live transcription (we are told Whisper or Deepgram), then an LLM agent that classifies intent and routes the request to the right team automatically: housekeeping, engineering, F&B, security. A guest saying "the AC in 412 is dead" should become a dispatched work order with no human in the loop. Accuracy has to be good enough in accented, noisy, multilingual environments (English, Arabic, Hindi, Tagalog) that staff trust it, or they will go back to the radio in a week. We also want a full-duplex conversational agent for the guest-facing line: barge-in support, multi-turn memory, live tool-calling into our PMS so it can actually check a booking rather than apologise. Sub-second round trip or it feels like an IVR. 2. Vision: make 400 cameras useful Real-time analytics over existing CCTV: crowd density in lobbies, loitering in service corridors, intrusion into restricted zones, unattended baggage, slip-and-fall detection. We need per-zone configurable thresholds, because a lobby that is normal with 40 people at check-in is an incident at 3am. False alarms are the entire ballgame here. A system that cries wolf gets muted by the duty manager on day three and we have wasted the year. Where feasible we want inference at the edge so we are not backhauling 400 video streams. 3. IoT: instrument the buildings We are deploying a sensor mesh: BLE staff badges (location plus panic button), ESP32 nodes for temperature, humidity, TVOC air quality, water-leak detection, door state, occupancy, and sub-metered energy. High-frequency telemetry over MQTT and BLE into a cloud pipeline, with dashboards that a chief engineer will actually look at. Some of this is genuinely high rate and lossy-tolerant, so we expect an opinion on UDP versus TCP ingest, on what belongs at the edge, and on how not to pay a fortune to store it all. 4. Knowledge: RAG over everything we have written down Twenty years of SOPs, brand standards, equipment manuals, supplier contracts, safety procedures, in PDFs nobody reads. We want staff and the voice agent to be able to ask questions and get an answer with a citation back to the exact page. Confidently wrong answers about a fire procedure are worse than no system at all, so retrieval quality, reranking, and traceability matter more to us than model choice. 5. Agentic operations: the actual product The layer that makes the rest worth buying. Work orders route themselves by skill, location, and current load. Escalation triggers when SLAs slip. Preventive maintenance is scheduled from sensor drift, not from a calendar. Alerts are contextualised against thresholds and history instead of firing raw. This is where we expect the most architectural judgement and the most pushback from you on what is realistic. 6. Predictive ML: stop reacting Forecasting on the telemetry and work-order history: chiller and HVAC failure prediction, occupancy and staffing forecasts, energy anomaly detection, laundry and linen demand. Plus OCR and NLP on supplier invoices and delivery notes so procurement stops being manual data entry. 7. Product surface: mobile and web Staff mobile app. Must work offline-first. Our engineering teams spend half their day in basements, plant rooms and lift shafts with zero signal. Local database as the source of truth, reconcile to cloud when the connection returns. Not a cache. We have been burned by this before. Web dashboards. Seven roles (GM, duty manager, chief engineer, housekeeping supervisor, security, F&B, corporate), each seeing only what they can act on. Real-time. Role-based access, SSO against our corporate identity provider, and a full audit trail, because this is guest data under regulation. 8. Platform: make it not fall over Migrate the existing Django and .NET Core services into a coherent event-driven architecture. Containerise, orchestrate, CI/CD, monitoring and alerting, cost control. We have 16 properties today and a mandate to reach 40. It must scale per property without our cloud bill scaling linearly with it, and a bad deploy must never take down a hotel's operations at 7pm on a Friday. Existing stack you would inherit (honest inventory) PMS: vendor product on Microsoft SQL Server, integrated via a .NET Core / C# middleware service a previous team wrote. It cannot be replaced this year. It has to be wrapped, not rewritten. Internal admin panel: Django, undocumented, one contractor, long gone. Guest-facing loyalty app: Firebase, works, do not touch it yet. Cloud: mostly AWS, some Azure for the corporate side. We are not religious about it. Everything else: greenfield, your call. Engagement structure We are not asking you to quote a full build blind, and we will not trust anyone who does. Phase 0, paid discovery (2 weeks). You audit what we have, sit with our chief engineer and a duty manager, and come back with a written architecture, a risk list, a build order, and an honest verdict on what is not worth building. If your verdict is that half of this should not be built, we want to hear that in week two, not month eight. Phase 0 is paid regardless of whether we proceed. Phase 1, pilot (roughly 3 months). One property, end to end. Voice dispatch, vision on the cameras in that building, the sensor mesh, the staff app, the dashboards. Real staff, real shifts, measured against the baseline you set in Phase 0. Phase 2, scale (ongoing). Roll to the remaining 15 properties, harden, hand off. Lead our two in-house engineers, help us hire two more, and own the technical story in front of our board and our next funding round. We are open to a fractional CTO title and equity conversation for the right person after Phase 1. Who we are looking for You have personally shipped a real-time AI system into production, not a demo, not a notebook, and you can talk about the part that broke. You are as comfortable arguing about MQTT quality of service and BLE power budgets as about retrieval strategy and eval sets. You have led a small team before. You can sit in a room with a hotel GM who does not care what a vector database is and come out with the right requirements anyway. Most importantly: you will tell us what not to build. We have already been sold an "AI platform" once by a team who agreed with everything we said. We are still paying for it. Deliverables Architecture document, threat model, and data-flow diagrams Working pilot in one live property, running on real staff for at least 30 days Voice, vision, IoT, RAG and agent services, containerised, with CI/CD and monitoring Staff mobile app (offline-first) and role-based web dashboards Evaluation harness and dashboards for model quality, plus the runbook that lets our team operate it without you Knowledge transfer to the two in-house engineers, and a hiring scorecard for the next two