Quirio · codename Liora
Production AI assistant for Texas residential real-estate agents. Listing discovery, comparative market analysis, client management, document drafting. Ran from 2023 to April 2026. Archived to focus on quant research. Post-mortem below.
What it was
An AI assistant for residential real-estate agents in Texas. Agents spend most of the day pulling listings for clients, building CMAs, drafting emails, and chasing leads. Quirio did all of that conversationally against the live MLS feed.
What we built
Listing discovery over the live Texas residential MLS — natural-language queries against the structured and unstructured fields ("three-bed under 600k in 78704 with a yard and not on a busy street"). A CMA generator that pulled sold comps weighted by recency, similarity, and proximity, and produced a draft narrative the agent edited rather than wrote. A lightweight conversational CRM ("leads I haven't followed up with in 10 days"). And a document drafter — offer letters, listing descriptions, follow-up emails — where Quirio drafted and the agent edited.
React + TypeScript front-end, Python backend, Claude and GPT in production for the conversational layer. Vercel for hosting. Real MLS data under the licensing constraints that implies.
What worked
Agents who used it kept using it. Early-access feedback was almost entirely about which features came next, not about whether it was useful.
I learned how to run LLMs as upstream dependencies — rate limits, structured-output validation, retries with backoff, golden-test sets for silent model swaps. All of that went into trading-algo and the risk controller in kite-algo later.
And three years of uptime against a real user base taught me everything I know about shipping. Deploys, regressions, support tickets, billing, MLS data licensing edge cases at 11pm on a Sunday.
What broke
Distribution. Real-estate agents are already served by an entrenched vendor ecosystem — MLS providers, CRM platforms, transaction-management tools — with incumbent contracts and trade relationships I had no way to compete with. Cold outbound was slow. I underestimated this by a multiple.
Timing. 2023 was early for "AI assistant for X" as a category that buyers evaluated. By 2026 it's a well-resourced, well-defined market, but in 2024 the people who'd have most benefited from Quirio didn't yet have the mental model to evaluate it. Being too early and being wrong look the same from the runway side.
Focus. Quirio plus trading-algo plus high school. The right call was to commit to one or wind Quirio down sooner. I split attention longer than I should have, and the marketing function suffered first.
Lessons
I built the product and didn't crack outbound. Two different jobs; I treated them as one.
By 2026 the moat is data access or distribution. Not the model. Vibe-coded "AI assistant for X" is now a category that anyone can ship in a weekend.
I wound it down deliberately — emailed users, archived the code, started this page. A slow fade would have been worse for everyone, me included.
LLM output is untrusted upstream data. The risk controller and audit logs in trading-algo and kite-algo are direct descendants of what Quirio taught me about composing untrusted model output with hard constraints.
Status
Archived April 2026. Site offline. Repo going public after a secrets/PII scrub.
mahimn · quirio · post-mortem · apr 2026