Sovereign Intelligence for Local Real Estate
In many cases, it already does. The asymmetry that matters isn't compute — it's information. CasaSol is the operational instantiation of that thesis, on a Mac Mini in the office and a MacBook Pro in the room with the client, in Marbella. Zero cloud. GDPR-by-design.
After The Sovereign Individual & Every Company Can Be a Palantir Now · Why CasaSol · We Didn't Notice · The Unfakeable Layer, localfirstai.eu
A frontier cloud model reasoning about a Marbella property has zero access to the agency's listing histories, viewer-aggregate patterns, comparable-sale deltas, and authored neighbourhood context. The local Coach has all of it.
This is the version of the claim that survives sceptical engineering scrutiny. "A smaller model wins" would be a weak claim. "A smaller model with full context beats a bigger model with no context" is the claim. The corollary the booth makes physical: a Mac Mini with all of an agency's lawfully-held context will out-answer Claude Opus 5 on questions about that agency's market, every time, because Claude Opus 5 cannot see the data.
General intelligence is commoditised at the API level. The portals own flat data — price, beds, baths, postcode. The only remaining moat for a Costa del Sol agency is hyper-local context: which streets flood, which schools have waitlists, which sub-areas turned over in Q1 and at what discount to asking. Today that context is wasted.
Most property data is the same property data. The villa on Idealista is the villa on the agency's own site is the villa in three aggregators' feeds — the same copy, re-scraped, often wrong, sometimes for a listing that sold months ago. Volume was never the problem. Trust is. On this coast one property is routinely listed by six agencies at six prices, and "available" frequently means "was." Scraping more of it produces more uncertainty at greater scale.
So CasaSol does not scrape. It witnesses.
Each listing enters as a photograph taken in person, on a given date, at a known location — a shopfront, a window, a printed sheet. The machine extracts the facts; the image stays as proof. The database does not hold "a listing found online." It holds "an offer observed to exist — here, then." That record cannot be reconstructed by a competitor with a scraper, because it was never on the wire. It is the first asset in this project that is verifiably ours.
The data is still thin, and it grows one street at a time. We are not pretending otherwise. We are choosing the slow, verifiable version over the fast, untrustworthy one — and making thinness the price of a corpus no aggregator can match.
CasaSol is the architect of the boundary between what the agency lawfully knows internally and what the cloud is allowed to see externally. The boundary is enforced by hardware and process permissions — not by prompt instructions.
Gemma 4 26B reads lawful agency data — listing histories, price-change patterns, days-on-market, aggregated viewer feedback, comparable-sale data, the agency's own neighbourhood guides. Emits agent_intel.json to the agency dashboard. Property and market patterns only — never inferences about identifiable natural persons.
Same Reducer, second pass, different system prompt. Generates public_listings.json against a versioned redaction schema enforced in code, not prompt. Inherits property-attribute tags from Tier 1 (commute time, school catchment, flood-extent overlap) — a richer surface than the agency could write by hand, with zero personal-data exposure.
Gemma 4 E4B fronts the MCP server, emitting routing plans against an ADR-001 discriminated-union schema. Plans dispatch to three tools — search_properties, get_property, market_summary — already proven against Claude Desktop. The MCP process has no read permission on the private layer. A prompt-injection attack fails at the OS, not the prompt.
The primary deliverable is the Property Intelligence Report — a structured PDF generated for any listing in under 60 seconds: deduplication across portals, price history, Catastro official sqm, SNCZI flood zone, beach distance, school proximity, and an LLM-generated assessment with a recommended offer range. One ref number in. One PDF out. Runs on the MacBook Pro in a client meeting.
The agency is sovereign over its lawfully-held listing data. The buyer is sovereign over their own query data, through their own contract with their AI provider. MCP is the protocol over which they transact. Neither surrenders data to the other; neither surrenders data to a third Leviathan.
Three places that hit the brief, with the worry attached to each.
Nueva Andalucía — three-bed townhouse, €385k. 42 min A-7 to PTA walkable · Aloha International outside 2024 SAIH flood layer Quiet cul-de-sac. Worry: ground-floor unit faces north, the courtyard sees light only in summer afternoons.
Cancelada — semi-detached with garden, €398k. 38 min A-7 to PTA walkable · St. George's Cancelada outside flood extent Worry: A-7 sound-bleed if the bedroom faces east; a viewing should happen during morning rush.
Sierra Blanca lower — two-bed apartment, €399k. 48 min A-7 to PTA walkable · Aloha Worry: at the upper end of your budget for the metres you get, and the comparable-sale layer shows the sub-area closed 11% under asking in Q1 2026 — there is room to negotiate.
Because the buyer is the controller of their query data and the agency is the controller of its listing data, the controller / processor mapping is unusually clean. This is the consequence of the architecture, not its purpose. The architecture is the sovereignty pair; clean compliance falls out of it.
The Cloud and AI Development Act defines four assurance levels for sovereign cloud computing services. Recital (10) of the Act explicitly removes "the AI system itself and its underlying model" from the scope of that definition. CasaSol delivers no cloud computing service. Your data runs on hardware you own, in a building you control, returning results that never cross a network boundary. There is nothing to certify — because there is nothing to regulate.
CADA's highest tier (Level 4) requires zero third-country control — a standard AWS, Azure, and Google cannot meet while subject to US law. OVHcloud and Hetzner reach Level 2 at best. CasaSol is structurally outside the entire framework: on-premise inference, no remote access model, no audit trail obligation because there is no network traversal to audit. The answer for any agency that asks: "CADA certifies cloud services. CasaSol has no cloud service. Your data never leaves your hardware. There is nothing to certify."
Spanish contracting authorities — Catastro, the Junta de Andalucía housing department, notarial chambers, municipal land registries — must procure cloud services at minimum Level 1 from 2028–2029, with public-order-relevant data at Level 2+. US hyperscalers are effectively disqualified from these categories. CasaSol is the architecture agencies adopt after conducting that sovereignty assessment: no cloud service to evaluate, no sovereignty level to certify.
CADA mandates that public bodies favour open-source and open-weight solutions over proprietary ones. CasaSol's entire inference stack qualifies: Gemma 4 (Apache 2.0), Ollama (MIT), SQLite (public domain), ChromaDB (Apache 2.0), mlx-whisper (MIT). A cloud AI competitor — Azure AI, AWS Bedrock, Google Vertex AI — fails the preference test by definition. Every component CasaSol runs is already CADA-aligned.
CasaSol ships with a compliance binder the agency hands to its DPO on day one. Eight artifacts, written by an EU GDPR practitioner, reviewed by the agency's own DPO before any corpus is generated. Bilingual (English canonical, Spanish executive summary) for OLÉ Marbella.
Pre-filled RoPA template, Article 30 aligned, ready for the agency to adapt and sign.
Vendor ↔ controller DPA between CasaSol and the agency.
LIA documenting the Article 6(1)(f) basis for the Coach layer.
For agencies surfacing Bouncer queries to public buyers.
Access, rectification, erasure, restriction, portability — with documented CLI commands for the local appliance.
72-hour template aligned with AEPD reporting requirements.
Tier 1 outputs retained on a documented schedule. Raw CRM never imported beyond the pipeline pass.
What the appliance logs, where, and for how long. Bounded to the physical machine.
Prefill cost on Gemma 4 26B-A4B grows super-quadratically past ~25K tokens of on-the-wire prompt. The bottleneck is memory bandwidth, not VRAM. Mitigations: cliff-aware coarsening, an 18K hard prompt budget, a streaming watchdog, and one-key restart of ollama serve.
Chronos Incident 003-Alpha · ADR-001 schema · OpenClaw cascade · Apple Silicon UMA
The agency's listing histories, viewer aggregates, comparable-sale deltas and authored neighbourhood guides are exactly the dataset that converts a generic LLM into local intelligence. None of it leaves the appliance. The buyer's Claude pattern means the cloud's discovery surface works for you, on your terms.
Public portals own static snapshots — beds, baths, asking price. True alpha lives in velocity. The v3.6 Temporal Delta Engine parses historical listing decay, viewing-frequency drops, and agent note sequences to surface real-time pricing leverage. Static facts belong to Idealista; market velocity belongs entirely to you.
moat economics · MCP integration · Temporal Delta Engine · vendor independence · what an alpha pilot looks like
Architecture forbids unlawful processing at the data-shape level, not at the policy level. Forbidden inputs and forbidden outputs are enforced in code via the redaction schema. Article 22 is out of scope by construction. The compliance pack — RoPA, DPA, LIA, DSR procedure, breach procedure — is part of the deliverable.
Static compliance binders are snapshots. The v3.6 Programmatic Compliance CI/CD layer sweeps every listing against 8 pre-registered GDPR liability categories before it reaches the public corpus — generating a daily record of zero leaks. The methodology is pre-registered and the evidence is public before the claim was made.
controller / processor mapping · lawful basis · CI/CD compliance gate · Exp 006 evidence · DPO walkthrough
Tacit knowledge does not extract itself. Agencies whose CRM is mostly contact data do not get a useful Coach. The asymmetry advantage requires data that is rich, consistent, and private. Where it isn't, no model — local or frontier — substitutes for the work of capturing it. Most booths claim everything; this one claims a limit, on purpose.
data-richness audit · what disqualifies a pilot · the Real Constraint · failure modes
No hidden infrastructure. The development substrate matches the deployment substrate — both inside the perimeter we ask buyers to trust.
The left screen runs two demos back to back. First: an agent voice note — recorded on a phone, AirDropped to the appliance — transcribed on-device by Whisper, split into private vault and public surface by the router, and redacted to a clean GDPR-safe listing in under 30 seconds. No cloud. No typing. Second: a synthetic toxic agent note redacted live by the same model, the same prompt that passed a pre-registered 20-fixture fidelity sweep with 0 leaks across 8 GDPR categories.
The right screen runs Claude Desktop against the live 70-listing corpus via MCP — real questions, real local data, real answers. Next to it: the B200 postcard, full-size, on heavy stock.
What you see was already running on that machine before the conference. The compliance evidence is pre-registered and public — run before the system was claimed to work.