Nidopr

Puerto Rico’s real-estate listings sit scattered across a dozen disconnected sources. There’s no single place to search the market, see prices, and trust what you’re looking at. Nidopr pulls that fragmented market into one bilingual (ES/EN), AI-priced, self-monitoring product. It runs in production on infrastructure I manage myself. I designed and delivered it while directing an AI-assisted implementation layer.

Problem

A buyer or analyst in PR has no equivalent of a national listings portal. The data is spread across 10+ sources. Formats disagree. The same property shows up in several places. And nothing tells you whether an asking price is reasonable. Nidopr is the missing data layer: acquisition, pricing, search, and the ops that keep it all alive, in one product.

Build: Data engineering

• 10+ acquisition sources unified into one schema, with resilient acquisition from difficult public sources (circuit breakers, resumable queues, graceful degradation).
• Cross-source entity resolution so the same property from different sources resolves to one listing, with four precision guards against false merges.
• Strict no-delete governance: soft-expire plus a nightly verify-or-expire pass, so the record stays honest and never silently drops data.
• Enrichment pipeline: regex plus local-LLM extraction, CRIM property-registry matching (~71.6%), geocoding, and photo junk-filtering, feeding a five-stage nightly quality pipeline.

Build: ML / AI

• LightGBM valuation model with split-conformal calibrated 80% prediction intervals, holding ~79.6% empirical coverage against the 80% target.
• A back-testing trust gate that cut model bias from +44% to +5%. It withholds low-confidence numbers instead of showing a figure it can’t stand behind.
• Hybrid semantic search: Postgres full-text plus pgvector/HNSW over 768-dimension embeddings, with graceful fallback.
• Local LLM pipelines (Ollama / qwen2.5) and a buyer-facing comparable-sales PDF report with an LLM narrative checked by a regex fact-checker.

Build: Product

• Nuxt 3 + TypeScript + Tailwind + Pinia + ECharts front end.
• An ES/EN natural-language search parser backed by a ~29k-check test suite.
• Auth with anonymous→login action-replay and Google OAuth; saved-search email alerts.
• Self-hosted behavioral analytics (funnels, rage-clicks, zero-result searches) and SEO with a ~22.4k-URL dynamic sitemap, verified in Search Console.

Ops

• Self-managed Ubuntu server, git push-to-deploy, Dockerized Postgres.
• 34 systemd timers running scheduled production jobs unattended nightly.
• Private networking via Cloudflare Tunnel (no exposed ports) and Tailscale.
• Restore-verified restic backups and an authorized security audit (21 findings, posture strong).

Sub-system: Nidopr Monitor

One pane of glass for the whole business. It alerts to your phone, with no vendor lock-in. An installable PWA (desktop and Galaxy Fold) running 40 production health checks across three alert sinks: Discord, email, and web-push to phone. It adds visitor analytics with bot/human/internal classification, a search→view→contact behavior funnel, a self-healing client, and a performance fix that took one endpoint from 4.8s to 0.1s (~48×).

Sub-system: Facebook auto-poster

A marketing channel that runs itself. The data finds the deal, writes the post, safety-checks the photo, publishes, and monitors its own health.

• Deals chosen by the data, reusing the site’s own below-market “heat” engine; rotates three categories (below-market sale / rent / Section-8) oldest-first with 30-day cooldowns.
• 7 posts/day at peak hours via the Meta Graph API, as proper timeline feed posts.
• An OCR photo-safety guard (tesseract): a phone-number watermark on any photo condemns the whole listing. It won’t funnel a caller to a competitor, and it won’t fall back to an unsafe image.
• Spanish captions generated from structured fields, plus self-monitoring. A Social dashboard tab logs every post and skip with a reason, backed by a shared invariant battery, schedule-aware liveness alerting, a kill switch, and token hygiene.

Numbers

• ~64,000 property listings aggregated and served (10+ sources, nightly).
• 640,000+ photos managed (~29 GB) with restore-verified backups.
• Valuation-model bias cut from +44% to +5%; ~79.6% coverage on calibrated 80% intervals.
• 40 automated health checks and 34 unattended nightly production jobs.
• 7/day autonomous social posts, deal-picked and OCR photo-safety checked.

Stack

Python · FastAPI · SQLAlchemy 2 · Postgres 16 · pgvector · Nuxt 3 · Vue 3 · TypeScript · Tailwind · LightGBM · conformal prediction · Ollama · tesseract · Docker · Cloudflare · Tailscale · restic.