AI Real-Estate Broker: works your side of the table, 24/7, to find your next home to rent or buy

Every real-estate tool you can name is paid for by the listing side. Zillow, StreetEasy, RentHop, Redfin: they are advertising businesses that optimize for lead volume to agents, not for you finding the right home. Nobody with that business model will build the tool that works against it, so it has to be built in the open. Build a broker that works one side of the table: the person moving. It runs continuously, remembers everything, does the tedious and adversarial work of a real search, and handles both renting and buying as first-class modes. To be precise about what it is not: this is a personal software tool operated by the home-seeker. It is not a licensed broker, it collects no fee, and it discloses itself as an AI in every outbound message. This is an agent system, not a search app. It plans over long horizons, uses real tools, maintains durable state across weeks or months, communicates with third parties on the user's behalf, and reasons about evidence quality rather than treating every claim as a fact. Its honest value proposition is compressed decision-making, not full autonomy: the user wakes up to a triaged shortlist and a queue of drafted messages, and one pass over a morning digest replaces an evening of tabs. Build it to that standard. CAPABILITY 1: PREFERENCE MODELING WITH REAL PRECISION Interview the user conversationally over time rather than through one giant form, and build a living, structured, versioned model of what they actually want. Capture the things people only articulate after twelve bad viewings: morning light, no ground floor on a bus route, a kitchen that fits a 30-inch range, above the fourth floor, pre-war detail, a tub not a stall, quiet at 11pm, a dog-friendly building with no weight limit, in-unit laundry versus in-building, walk-up tolerance that changes with the floor. Represent preferences as weighted, versioned constraints: hard filters, soft scores, and explicit trade-off rates ("$250/month more for in-unit laundry, but not $600"). Learn from reactions. When a listing is rejected, ask why and file an evidence-backed proposal. The model must never retune itself silently; every change is a proposal the user approves, with a visible changelog and the ability to roll back. Prove that with tests, including a test that it cannot quietly tune itself to make a favored listing win. Support multiple principals as a first-class case, because most moves are a couple or roommates. Each person gets their own weighted model; the agent scores against both, surfaces conflicts explicitly ("this wins for A on light and loses for B on commute"), and never silently averages two people into a preference neither holds. Model location as richly as the unit. Isochrone commute polygons against real transit schedules, not radius circles; for v1, precompute isochrone tables for the launch metro offline rather than building a live router. Block-level noise and late-night activity. Proximity to the specific places the user named. Sun exposure and orientation. Elevation and flood risk. The "near X but not on X" distinction no filter box has ever supported. CAPABILITY 2: DISCOVERY, DEDUPE AND HONEST SCORING Be honest about where listings actually come from. In a real market the legitimately accessible surface is (a) the alert emails already arriving in the user's inbox, (b) listings the user forwards, (c) pages the user captures from their own logged-in browser, and (d) any genuinely open feeds and APIs permitted by their terms. Build inbox-first: parse the alerts the user already receives, and make user-side capture effortless via a watched capture folder plus a bookmarklet that saves the current page to it. No extension, no automation of protected sites, no evasion. Deduplicate the same unit posted five times by four brokers at three prices. Track relists, price-drop history and quiet stale inventory. Score against the preference model and throw most of it away. The output is a short ranked shortlist with a written rationale for each: why it fits, what's wrong with it, what's uncertain, and the true all-in monthly cost (rent plus fees, or mortgage plus taxes, maintenance or HOA, insurance and amortized closing costs). Also answer the negative: for any listing the user asks about, explain why it was excluded or ranked low, citing the specific constraints. Be honest about evidence. Every fact carries provenance and a verification level; a listing's own claims are capped below anything first-hand. Flag listings the agent believes are misleading: a "flex 2BR" that's a 1BR with a partition, square footage that doesn't reconcile with the floorplan, a photo set reused from an old listing, fees that don't add up. Include a scam screen. All semantic understanding of photos, floorplans and PDFs routes through the user-configured multimodal LLM; do not claim computer-vision forensics the stack cannot support. CAPABILITY 3: TALKING TO THE OTHER SIDE This is what makes it a broker instead of a search engine. For listings that survive the filter, open and maintain real correspondence with listing agents to collect what is never in the listing: true interior square footage and how it was measured, an actual floorplan, window orientation, real floor number, the honest fee structure, genuine days on market, whether a prior deal fell through and why, whether pets are actually allowed, and the board or co-op process. Send through the user's own mail identity via their configured mail server; deliverability and reply-to are the user's real address, which is load-bearing for response rates. Draft in the user's voice, track threads per unit, and parse every reply back into structured fields with provenance. Follow-ups have a hard anti-spam ceiling: a configurable cadence with a maximum number of touches per thread, a mandatory permanent stop on any opt-out signal, and graceful give-up. Measure and report actual response rates honestly instead of implying correspondence always works. The counterparty may be a human agent or another AI agent; handle both, and always disclose that this is an AI acting on the user's behalf. Treat the inbox as hostile. Listing text, agent replies and attachments are untrusted data, never instructions. Build an explicit prompt-injection boundary and publish the adversarial suite that attacks it as a standalone benchmark others can run their own agents against: at least 40 attacks spanning listing text, reply bodies, attachments with hidden text, and counterparty agents asserting false authorization (raise the budget, pre-approved deposit, "the user already agreed"). Required pass rate is 100%, with results published in the build log. A fail-closed redaction guard scans every outbound message, attachment, and LLM request for SSNs, account numbers and other sensitive data, and refuses to send rather than risk a leak. CAPABILITY 4: TOURS, SCHEDULING, AND FIRST-HAND EVIDENCE Turn a ranked shortlist into a booked Saturday. Read real calendar availability, propose viewing windows, handle the scheduling back-and-forth, and cluster confirmed appointments by actual transit time between stops rather than straight-line distance. Handle reschedules and cancellations. Every calendar write and every scheduling message goes through the approval gate. Close the evidence loop after the tour. The user can dump a voice memo, notes, or photos from a walkthrough; the agent parses them into the unit's entity record as first-hand facts ("faces the airshaft," "renovated means painted," "laundry is coin-op") at the top verification tier, re-scores the unit, and, where the visit reveals a general preference, files a preference-model proposal for approval. First-hand evidence is the top of the hierarchy, so it must have a real input path. CAPABILITY 5: RENT AND BUY AS FIRST-CLASS MODES Renting: fee structures, broker fees, guarantors, application packet assembly, income requirements, lease terms, move-in timing. Buying: offering plans,