Skip to content
    Case Study / Operations AI

    Aetherix

    The F&B execution agent for boutique hotels.

    Tomorrow's covers, a staffing grid that fits, and less waste. One message a day, in WhatsApp, with the reasoning attached.

    Aetherix is the food and beverage agent of a five-agent architecture. The architecture

    Read withChatGPTClaude

    • Phase 3 · Pilot partner in recruitment, go-live September
    • Staging Live
    • MCP-Native
    • Claude Sonnet
    • Python 3.12
    • FastAPI
    • Fly.io (Paris)

    What it already does.

    • States its drivers the evening before, then reports its own miss the next morning, in plain language.
    • Flagged a corrupted POS export rather than training on it.
    • Named its own drift after three consecutive misses, and recovered from a 25% regime shift through weekly recalibration.
    • Stores every recommendation next to what actually happened, with the manager's override recorded beside it.
    • Blocks its own merges on evaluation exit codes; every guardrail trip carries a machine-readable reason.

    Sandbox data, real mechanics, deterministic reruns.

    Target (Design): overrides recorded with the responder's role and reason, not a named person.

    Three promises, one message a day.

    Two of the three are pilot hypotheses, not measured results. They are the outcomes the live pilot is set up to test.

    Running daily

    Tomorrow's covers

    Running daily against real data. Operator delivery begins with the pilot. Anticipates how many breakfasts, lunches, dinners to expect, with a confidence score and the drivers behind it.

    Pilot hypothesis

    Staffing that fits

    Proposes a shift grid based on occupancy, capture rate history and local context. The manager approves, adjusts or overrides.

    Target outcome

    Less waste

    Spots over-ordering risk before purchase, with the historical signals that triggered the flag.

    Understand, measure, learn.

    Understand, like a human.

    Contextual reasoning grounded in each property's history.

    Measure, like a machine.

    Every recommendation is stored next to its real outcome, so the system knows what it said and what happened.

    Learn, like a network.

    Per-property memory today, federated priors: Research, to give independents the power of a hive.

    Hospitality Multi-Agent Architecture

    Five agents, honest status.

    • AetherixThe F&B agent: forecast, staffing, memory.Built
    • TacetThe environmental-signal agent.Built · maintenance mode since July 2026
    • AnimaThe guest memory node.Synthetic PoC
    • PeritiaThe trade agent: house knowledge, what the people who work here know.Design
    • OrchestratorThe supervisory agent.Design · proto-stub

    Where this sits

    • AetherixYou are here

      A food & beverage execution agent for hotels. It tells a manager how many covers to expect, how to staff for them, and how to cut food waste.

      For
      Other agents via MCP, or a hotel / F&B manager with WhatsApp.
      Status
      Built, 0 real users, forecast in shadow mode, pilot partner in recruitment.
    • Headless environmental twin: street, weather and crowd signals turned into yield rules, exposed via MCP.

      For
      RMS, PMS and orchestration agents, not humans.
      Status
      Built, maintenance mode since July 2026.
    • A voice AI mentor that captures, qualifies, and leverages the tacit expertise from senior to junior collaborators. Just ask.

      For
      A junior technician on the shop floor, hands busy.
      Status
      Prototype, synthetic data, harness in CI.
    • Guest memory node: it keeps what a guest said and what happened, so nobody has to ask twice, without watching.

      For
      The nodes that need to know who is arriving.
      Status
      Synthetic PoC, production gated behind a data-protection assessment.
    • Open-data SaaS mapping urban noise so cities can act on acoustic health.

      For
      European cities and metropolitan areas.
      Status
      Shipped 2020 to 2022, company closed.

    How these systems fit together: the architecture note

    Project Info

    Role

    Product Builder & AI Agent Designer (solo)

    Duration

    2026, ongoing (validation phase)

    Team

    Solo architect & builder

    Repository

    GitHub Repository

    Public meta-repo for the Hospitality Multi-Agent Architecture

    Status

    Built · 0 real users · forecast in shadow mode · pilot partner in recruitment · Apaleo OAuth + MCP read operational

    Closed-loop demo, coming soon

    A three-minute walk-through of the loop: signal in, forecast, WhatsApp receipt, manager reply, memory update.

    Why it exists

    Your greatest operational asset walks out the door.

    In hospitality, the people who know how the property really runs (which Tuesdays spike, how a local conference reshapes breakfast, how to staff a rainy Sunday) carry that knowledge in their heads. When a manager leaves, the property starts again from scratch.

    Combined with the inherent complexity of operations (multiple outlets, mixed guest segments, volatile demand, labor shortages), this creates chronic inefficiencies:

    • Persistent over or under-staffing
    • Avoidable food waste
    • Reactive decision-making instead of anticipation
    • Slow onboarding and inconsistent execution across properties

    Traditional systems store data. Very few build living operational memory that learns from real outcomes and makes that knowledge available at the exact moment it is needed.

    This is the gap Aetherix was created to close.

    The underlying bottleneck is not data, it is synthesis: combining internal context (occupancy, capture rates, POS history) with external signals (weather, local events) into a recommendation calibrated to this property.

    The cycle

    Memory, Prediction, Action, Learning.

    Memory is not the offer, it is the mechanism that makes the forecast specific to one property rather than generic. Every day, Aetherix runs the same loop and adds one more record: recommendation, manager reply, measured outcome. This is why the day-90 recommendation is expected to be sharper than the day-7 one.

    01

    Memory

    Recall what worked here, on similar days, in similar contexts. Local history first, plus generic cross-property context patterns while local history is thin.

    Day 7Cold-start. Confidence is honest about what is not yet known.

    Day 90Expected: drivers become property-specific, calibrated by enough Tuesdays. This is what the pilot measures.

    02

    Prediction

    A calibrated forecast for tomorrow: covers, staffing, waste risk, with the drivers behind it.

    Day 7Baseline leans on generic patterns and first-week local signals.

    Day 90Expected: tighter confidence intervals and fewer overrides, to be verified on live data.

    03

    Action

    One message on WhatsApp. Reply accept, reject, or the right figure. No new dashboard to learn.

    Day 7The receipt is verbose, spells out every assumption.

    Day 90Intended: compact. The manager reads a paragraph, replies in a word.

    04

    Learning

    An exact accept or reject, or a corrected figure, is recorded as feedback; free-text replies are not yet mapped. The pilot tests whether feedback improves future recommendations.

    Day 7Every recorded accept, reject or correction deposits one grain of memory.

    Day 90Hypothesis under test: the system reflects how this property behaves, not just how hotels behave.

    Caveat: what is recallable today is local history plus generic cross-property context patterns. A true federated cohort prior (HiveMemory) is research, targeted for Phase 4 and beyond, and would remain anonymized aggregates only. No cross-tenant raw data, ever. Measured gains begin with the live pilot, not the synthetic baseline.

    Daily Receipt

    What a manager actually sees.

    Aetherix

    today · 18:42

    Daily Receipt
    Breakfast    84 covers   ·  confidence 88%        (illustrative)
                 72 rooms booked + 12 typical Tue walk-ins
                 Staff: 4 FOH  (you ran 3 last Tue → 1 understaffing flag)
    Driver       local conference check-ins + clear weather
    Last week    forecast 79 · actual 81 · MAPE 2.5%

    One reply: accept, reject, or the right figure. That reply is the signal Aetherix learns from. The pilot tests whether the day-90 receipt is sharper than the day-7 one.

    Anticipation, inspected
    Forecast
    74
    Actual
    71
    Variance
    -3 · 4%

    DecisionPrep cut on two slow-moving dishes, one runner moved to the later shift.

    ReadingForecast held. Waste stayed inside the usual range, no manual override needed.

    Illustrative figures from shadow-mode runs, not a signed pilot

    The architecture, bounded context by design

    Aetherix executes. It does not orchestrate.

    Aetherix is one node in a network of specialized, -native agents. Each node owns a strict bounded context and communicates only through MCP, no shared code, no monolith to displace. Target: the Orchestrator (Design) decides when to call it. Until it is built, Tacet's output reaches Aetherix directly: the one documented exception, non-personal, read-only, logged.

    Tacet

    Sensory

    Orchestrator

    Supervisory

    Anima

    Guest memory

    Synthetic PoC

    Aetherix

    F&B execution

    this node

    HiveMemory

    Anonymized peers, side input

    Manager

    WhatsApp

    Aetherix executes. It does not orchestrate.

    Target: the Orchestrator (Design) calls both. Until it is built, Tacet's output reaches Aetherix directly: the one documented exception, non-personal, read-only, logged.

    NodeRoleStatus
    AetherixF&B execution: forecast, staffing, memory, receiptsLive (staging)
    TacetSensory: weather, noise, open data → riskBuilt · maintenance mode since July 2026
    PeritiaHouse knowledge: what the people who work in a property knowDesign
    OrchestratorSupervisory: MCP routing, business rules. Proto-stub in Phase 3, dedicated repo in Phase 4.Design · proto-stub
    AnimaGuest memory nodeSynthetic PoC, hackathon prototype; production DPIA-gated
    Architecture

    Push-first pipeline.

    Apaleo (OAuth2 and MCP read) + Open-Meteo + PredictHQ → Prophet forecast and staffing → pgvector memory recall (OperationalMemory + HiveMemory cold-start, research) → Claude Sonnet reasoning → WhatsApp or Email.

    Context inputsFat backend, FastAPIIntelligence layers
    Internal context

    Apaleo PMS

    OAuth, MCP read

    External context

    Weather

    Open Meteo

    Events

    PredictHQ

    Social sentiment

    Web, reviews

    PMS Sync

    Capture rate

    Prophet

    Forecast & staffing

    Claude Sonnet

    Reasoning, fallback Gemini/GPT

    Explainability

    Why? service

    Supabase

    Operational data

    pgvector

    Per-hotel memory

    HiveMemory

    Anonymized peers

    LLMProvider

    Claude Sonnet · Gemini · GPT

    MCP server

    Three atomic tools, multi-tenant

    Three atomic tools
    WhatsApp push
    One-tap human approval

    Numerical · Semantic · Cognitive · Reasoning · Action

    Stack

    • FastAPI
    • Python 3.12
    • Prophet
    • pgvector (Mistral 1024d, HNSW)
    • Claude Sonnet (Gemini/GPT fallback)
    • Redis (Upstash)
    • Supabase (PostgreSQL, RLS, EU)
    • Fly.io (Paris)

    Trust is engineered, not promised.

    Evaluation as a contract

    Qonto-inspired. Versioned golden datasets. Any change to forecasting, providers, memory or prompts needs a green eval pipeline to merge. A hard CI gate, not a habit.

    Observability & metering

    Every MCP tool call is one structured event. Auditable, queryable, priceable. The product is operable, not opaque.

    Multi-tenant security

    hotel_id scoped at every layer (RLS, Python guards, MCP token binding), resolved server-side, never a tool input. Read-only PMS access. GDPR: Supabase EU, HiveMemory is anonymized aggregates only.

    Trust engineering, visible

    Three artifacts, three proofs.

    Screenshot coming
    Eval pipeline, hard CI gate
    Screenshot coming
    MCP tool-call observability event
    Screenshot coming
    Multi-tenant hotel_id scoping

    What remains private

    The project includes proprietary research on cross-domain operational context and cold-start learning. Public materials focus on the engineering evidence: evaluation gates, traceability, safety guardrails, privacy boundaries, and outcome measurement.

    • Evaluation gates
    • Traceability
    • Safety guardrails
    • Privacy boundaries
    • Outcome measurement

    Why Lore cannot close the same loop

    Lore, in aviation maintenance, shares this architecture and cannot close the same loop. Nothing there tells the system whether an observation was right, and the moment a technician escalates the counterfactual disappears. Aetherix has an outcome to compare against; that difference decides how much each product is allowed to learn on its own.

    The evaluation vocabulary behind these gates, harnesses, graders, acceptance tiers and closed loops, is written up in a field note: Harnais, graders, boucles fermées (in French).

    Where we are, honestly.

    No live forecasting accuracy claim, yet. Phase status, target outcomes, and the security gate baseline. Every claim carries its caveat: synthetic, target, or measured.

    Fig. 9.1 · Anticipation vernierJ+1 · synthetic backtest
    0%5%10%15%20%25%30%FORECAST ERROR · MAPELOWER IS BETTERGATE · 22%15.09%CALIBRATION BASELINEUNREAD
    EvidenceSynthetic backtest
    Security14/14 gates passing
    Reality check0 real users
    Instrument at rest
    Synthetic calibration baseline, not a live accuracy claim. Real-world measurement begins with the pilot.

    Where we are, where we're going.

    1. Phase 0

      Validation

    2. Phase 1

      Observability

    3. Phase 2

      Receipts

    4. Phase 3

      Pilot

      Pilot design partner in recruitment. Go-live targeted end of September.

      Current
    5. Phase 4

      Scale

      Mews · POS · Federated HiveMemory · Context Graph · Anima

    6. Phase 5

      Vision

      Revenue Orchestrator

    Sister node · Guest Memory

    Anima. Guest memory, four layers deep.

    Where Aetherix remembers the property, Anima remembers the guest. A synthetic proof of concept built at the Anthropic hackathon (June 2026): a four-layer temporal memory with confidence-weighted claims, run on synthetic data. Production is Phase 4, DPIA-gated.

    Explore the architecture
    1. Layer 01

      Working

      The current stay, in progress. Expires at check-out.

    2. Layer 02

      Episodic

      This stay and a short tail after it. Weeks.

    3. Layer 03

      Semantic

      Durable preferences, stated or repeatedly confirmed. Long, with decay.

    4. Layer 04

      Segment

      Anonymised patterns, what guests like this tend to need. Aggregate, never individual.

    Role & pricing

    Solo architect & builder. Results-as-a-Service.

    Product Builder & AI Agent Designer, 2026, ongoing (validation phase). Pricing follows outcomes, not seats: we win when you win. Pilots and product conversations open.

    Thank you for your time.

    In a crowded field, sustained attention is the rarest courtesy. I hope this work offered both enlightenment and quiet clarity.

    Conceived and crafted in Paris · Built 0→1 · Verified through open specifications

    © 2026 Ivan de Murard · Curated with Care