On this page
Case study · AetherixPre-pilot
The F&B agent for boutique hotels.
It works inside the tools the hotel already runs, its PMS and WhatsApp: one recommendation a day, the manager's reply, the outcome recorded next to it. Built to improve with every day it runs.
- Role
- AI Product Manager and Product Builder, solo · 2026
- Status
- Built and tested on the Apaleo sandbox and on public restaurant data. 0 real users. Being made pilot-ready; no hotel yet.
Covers expected: 101 Why 101: usual level 87; hotel 98% full +9, New Year's Day +8, season +7, weather −6; other factors −4. High confidence: 94–109 covers on 4 nights out of 5. Staffing: 7 servers, 2 hosts, 3 runners, 5 kitchen staff, on the hotel's own rules (usual night: 5, 2, 2, 3).
Why 101?
- Usual level
- 87
- Hotel 98% full
- +9
- New Year's Day
- +8
- Season
- +7
- Poor weather
- −6
- Thursday
- −3
- No local event
- −1
- Forecast
- 101
The range is the model's 80% interval. It is 15 covers wide, 15% of the forecast. Under 20% earns the engine's top rating, 0.9. The rating measures how tight the range is; it is not a probability.
The hotel's own rules: 1 server per 15 covers, 1 host per 60 covers, 1 runner per 45 covers, 1 kitchen staff member per 25 covers, rounded up.
I'd say 79 covers instead.
modifiedThe simulated manager averages the last 4 Thursdays: 69, 81, 72 and 95.
Actual: 95 covers. Forecast 101: 6% off, inside the 94–109 range. Your estimate 79: 17% off.
TL;DR
- The problem
- What a hotel knows about its own service lives in a manager's head, and leaves with them.
- The bet
- One recommendation a day in WhatsApp, checked against what happened the next morning.
- Today
- The loop runs daily on Maison Lumen, a fictional hotel in the Apaleo sandbox. The product is being made ready for a first pilot; no hotel yet.
- Still open
- The forecast is not yet proven on hotels. No pilot hotel has signed.
- Built
- Implemented behaviour that can be inspected in the product or repository.
- Validation
- A bounded test, audit or measurement. It does not prove every outcome.
- Design
- A proposed direction or a prototype.
01 · The problem
The knowledge leaves with the manager
Tomorrow's covers and staffing depend on what one person remembers: last year's conference week, the regulars, the weather. The result is over or under-staffing, avoidable waste and decisions made on the day.
The data exists. What is missing is the synthesis: occupancy, sales history, weather and local events turned into one recommendation for this property.
Where this comes from. Informal conversations with product managers who built staffing and scheduling software for restaurants and hotels. No F&B manager has been interviewed yet; that is part of the pilot.
02 · How it worksBuilt
One record a day
What was recommended, what the manager said, what happened. The message in the demo above is real sandbox output.
1 · Evening
Tomorrow's covers and staffing, with the reasons.
2 · Reply
The manager accepts, rejects, or gives the right figure.
3 · Next morning
The real outcome is stored next to the recommendation.
4 · Adjust
The agent recalibrates on the gap.
Free-text replies are not used yet. Whether the agent gets sharper over time is what a pilot would test.
03 · Product strategyDesign
Understand a property the way its manager does
And measure every recommendation against what happened. Five decisions follow from that.
- 1
Distribution before design
A hotel floor has no time for a new tool. So the product plugs into the PMS and is built to be listed where hotels already pick their integrations.
- 2
F&B first
A covers forecast has a result the next morning. It can be checked.
- 3
WhatsApp as the only surface
It shows the recommendation, keeps a human line to the manager, and records the reply the agent learns from.
- 4
Three parties: who uses, who pays, who distributes
The F&B manager uses it. Hotel management pays. The PMS distributes.
- 5
Pricing on results, to start
With no track record, the builder carries the adoption risk. The later model is undecided.
Distributes
The PMS
Apaleo shares read-only data and lists the integration.
Aetherix
Uses
The F&B manager
Gets the daily message. Sends the reply.
Pays
Hotel management
Signs the contract.
Unproven
- No hotel or PMS has validated this route yet.
- The platform could build the same thing.
- Pricing on labor cost needs payroll data the PMS does not hold.
04 · What is proven, what is notValidation
The loop is built. The forecast is not yet proven on hotels.
Already working
- States its drivers the evening before, then reports its own miss the next morning.
- Rejected an implausible POS figure (850 covers, injected as a test) instead of learning from it.
- When recent data diverges sharply from history, says so and asks the manager to double-check.
- After a simulated 25% jump in demand, weekly recalibration cut the forecast error from 15.9% to 12.3%.
- Stores every recommendation next to what happened, with the manager's override beside it.
Sandbox data, real mechanics, deterministic reruns.
The forecast test
Public data, 30 restaurants. Against a naive same-weekday baseline (the mean of the last four same weekdays): mean error 49.71% vs 55.86%, ahead on average. For next-day forecasts, the horizon the product uses, 28.5% vs 27.9%, slightly behind.
The test used restaurant data, without hotel occupancy, the signal the product relies on. The protocol was not pre-registered. Maison Lumen uses occupancy from Apaleo; how accurate the forecast is on a real hotel is what a pilot will measure.
Kaggle Recruit, 30 Japanese restaurants out of 829, 1,193 scored points, deterministic rerun. Prophet against a naive same-weekday baseline (the mean of the last four same weekdays): mean MAPE 49.71% vs 55.86%, passing the pre-set 5-point rule. MdAPE 27.0% vs 27.6%. At J+1, 28.5% vs 27.9%. A gradient-boosted model reaches 25.6% on the median. The protocol was not pre-registered for this run.
MAPE. Mean absolute percentage error. The average gap between a forecast and what actually happened, expressed as a percentage, which makes errors comparable across covers, revenue, or volumes.
Phase 0–2
Validation, observability and the Daily Receipt are done.
Phase 3 · now
Pilot readiness. No hotel yet.
Phase 4
Scale: Mews, POS, Anima.
Phase 5
Revenue Orchestrator. Hive Memory stays research.
05 · The Decision LedgerBuilt in AetherixDesign as a shared layer
The outcomes layer the others do not have
Every recommendation Aetherix sends becomes one row: the context, what it recommended, what the manager said, what happened. That ledger is built, with no real user yet. The bet is to make it a layer: every agent of the architecture writes its rows there, and agents built by others read it through MCP.
The ledger, one row per decision
| Agent | Recommended | Human said | What happened | Gap |
|---|---|---|---|---|
| Aetherix · built | 101 | 79 | 95 | 6% vs 17% |
| Tacet, Anima, Peritia | Same row, written by every agent of the architecture · designed | |||
| Agents built by others | Read and written through MCP · designed | |||
First row: built, shown on the sandbox run at the top of the page. The other rows are design.
Why it matters
Agents on marketplaces such as Apaleo's Agent Hub recommend. The ledger is the part that records what happened next, for Aetherix's own agents and for others'.
Built in Aetherix on 1 August 2026, no real user yet. The shared layer is design.
The shared layer: design, nothing built yet
- The ledger has no real user yet, and Aetherix no pilot hotel.
- Other agents would need consent, a shared format and access to outcomes.
- The platform could build the same layer.
06 · How it's builtBuilt
Specified, reviewed and tested before anything ships
- Spec before code, with the BMAD method. Requirements, architecture, epics, then one written story per change. Code comes last.
- I direct the agents; I own the result. AI agents build from those stories, inside 34 decisions I set. I review and decide what ships.
- Nothing merges without passing its tests. A fixed set of scenarios checks every change; a drop in quality blocks it.
- A weekly self-audit. An agent checks that what the project claims still matches what it does.
Inside the agent
One pass a day, on Maison Lumen's first evening: the same figures as the demo at the top. Pick a step.
Manager
I'd say 79 covers instead.
Aetherix
Actual: 95. Forecast 101: 6% off, inside the range. Your estimate: 17% off.
Outcome recorded
The reply and the real outcome sit next to the recommendation. That record is the ledger.
Around every step: evaluation gates that block a change, one hotel's data never visible to another, data kept in the EU.
Where it sits
Aetherix is one agent of the Hospitality Multi-Agent Architecture. It executes F&B; it does not orchestrate. Every exchange goes through MCP.
Orchestrator
Routing and business rules
Design, proto-stub
Aetherixthis page
F&B execution
Built · forecast in shadow mode
Tacet
Sensory: weather, noise, open data
Built · maintenance
Anima
Guest memory
Synthetic PoC · DPIA-gated
Peritia
House knowledge
Design · pattern tested in Lore
Decision Ledger · built in Aetherix; shared across agents: Design
Tacet → Aetherix: direct, until the orchestrator exists
Third-party agents read it, through MCP · designed
Pipeline and stack
Apaleo (OAuth2 and MCP read) + Open-Meteo + PredictHQ → Prophet forecast and staffing → pgvector memory recall (Operational Memory + Cold-start prior, research) → Claude Sonnet reasoning → WhatsApp or Email.
Apaleo PMS
OAuth, MCP read
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
Hive Memory
Anonymized peers
LLMProvider
Claude Sonnet · Gemini · GPT
MCP server
Three atomic tools, multi-tenant
Numerical · Semantic · Cognitive · Reasoning · Action
- FastAPI
- Python 3.12
- Prophet
- pgvector (Mistral 1024d, HNSW)
- Claude Sonnet (Gemini/GPT fallback)
- Redis (Upstash)
- Supabase (PostgreSQL, RLS, EU)
- Fly.io (Paris)
Built to be trusted
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, Hive Memory is anonymized aggregates only.
More
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
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 in What is an LLM evaluation harness? Graders, protocols and closed loops for AI agents
Aetherix is one of five agents that talk to each other only through . The architecture note says which are built and which are design. Anima, the guest memory node, has its own case study.
Next · where Aetherix sits
Hospitality Multi-Agent Architecture
Five specialised agents, each owning one bounded context, talking only through MCP. The page says which is built and which is design.
Aetherix, Tacet, Anima, Peritia and the orchestrator
Read the architecture07 · Talk
About a pilot, or a role
Open to a first pilot hotel running Apaleo, and to AI Product Manager and Product Builder roles.