BUILT BY ONE OPERATOR.
1.5B+ RECORDS.

The expansion engine joins TX OSSF (On-Site Sewage Facility) permits with subdivision polygons across 8 counties (Travis, Williamson, Bastrop, Hays, Comal, Bexar, Guadalupe, Kendall) to identify dense septic clusters. 690 high-confidence septic streets were extracted from the existing Mac Septic customer base as ground truth.

A pilot of $4,200 BatchData spend on Cedar Creek + East Austin (42,539 addresses) is projected to yield 150-200 net-new MAC customers, based on the historical cluster density observed at Summit Glory Trail (307 Mac customers in one subdivision). Enrichment is the gating step: 107,562 candidates remain unenriched and will be processed in cost-managed batches.

The 2026-05-16 acquisition of the FCC Broadband Data Collection (924M records, 116M unique US addresses) added a free pre-scoring layer. Every TX OSSF permit candidate now carries a rural_septic_score 0-100 derived from broadband absence + TIGER urban-area overlap + ZCTA population density. Mac Septic's 8,630 customer addresses validate the v2 model — they hit the 70+ score band at 4.19% vs 1.84% for a random TX baseline (2.29x lift). BatchData skip-trace is now reserved for confirmation on the top-ranked 4,190 addresses, not blanket enrichment of all 42K.

The public face of this data layer is the National Septic Systems Map (linked above): 17M+ septic systems rendered on an adaptive cluster grid served straight from the PermitLookup API — Florida confirmed from state DOH permit records, other states inferred from property and permit signals, with the confidence tier disclosed per state on the map itself.

A 17.3M-row materialized view joins NOAA NCEI storm reports against permit data across TX. Validation cases: Tarrant Co May 27 2024 hail (3.25") drove 1,006 roof permits in the following 120 days; Bexar Co Mar 31 2025 hail (2") drove 3,406 permits. The product surfaces address-level lists pre-filtered by storm severity and roof age.

The data layers (austin_311, storm_events, spc_storm_reports, hail_leads MV) are live; the remaining work is a paid `/api/v1/hail_leads/search` endpoint with Stripe gating and a refresh cron. Real-world launch is imminent — the lead surface is live in the permitlookup-web Next.js app at hail.ecbtx.com, gated and ready for first pilot traffic.

Rural Service Lead Engine v5 is a scoring layer over a 4.25-billion-row federal-and-state cluster: 924M FCC broadband records, 1.75M parcel polygons (HCAD 1.46M Harris + Comal 102K + Hays 119K + Bandera 32K + Kerr 34K = full Texas Hill Country + Houston coverage), 2,644 TIGER urban-area polygons, 33K census ZCTAs, and 228K TX OSSF septic permits. For every address it computes a rural_septic_score 0-100 combining: TIGER urban-area NOT-overlap (anti-signal -35), polygon-derived lot_acres (primary signal up to +35), population density, broadband absence, and joint-condition interactions.

The validation story: v1 used broadband absence alone (0.29x lift — failed). v2 added urban-area + density (2.29x lift — usable). v3 tried address-based parcel join (1.56x — regressed due to 0.012% match rate). v4 used HCAD spatial join (0.43x — inverted because Mac is in Hill Country, not Harris). v5 added Hill Country parcel polygons for symmetric enrichment of both cohorts (5.99x vs all-TX, 5.91x vs HGAC — SHIP). Logistic regression AUC 0.86, lift 4.09x independently confirms — no leakage.

Output: 4,145 Mac-customer-confirmed rural targets in the dialer-ready queue, plus 10,000 ranked TX OSSF permits as expansion candidates. Symmetric spatial enrichment is the key — both cohorts get apples-to-apples treatment so the lift number is real, not an artifact.

`/v1/roofer-leads/by-hail-event?event_id=X&days_after=120&radius_miles=20&min_magnitude=1.0` returns properties within a hail event's footprint, ranked by composite score: storm_severity (hail magnitude × 10, capped 30) + home_age (years / 2, capped 25) + mortgage_score (+20 if HMDA match within window) − roof_recency_penalty (−20 if a roof permit was pulled <5y ago, −10 if <10y) × distance_falloff.

`/v1/roofer-leads/recent?state=TX&days_back=14&min_score=50&limit=100` runs the same scoring against all hail events from the last 14 days in a state.

The underlying data was acquired and indexed during the 2026-05-15 federal-bulk session: 771K NOAA storm events 2014-2025 (771,183 details + 690K locations + 11K fatalities), 35M HMDA mortgage applications 2020-2024 (every covered lender, every loan purpose, every action), 97M property sales, 297M code violations. The state-partitioned permit tables filter recent roof permits per state.

Delivery: free for existing PermitLookup customers; standalone pilot at $50-100/lead × 100 leads/mo per pilot customer.

As of 2026-05-17, Storm Strike Dispatch ships as a focused lead-acquisition product at storms.ecbtx.com. Buyers who want a full operating system (CRM + dispatch + estimates + claims) are routed to its sibling product RoofCRM at roofers.ecbtx.com — same data layer, different value prop.

Three FastAPI endpoints live on production (Railway + Cloudflare Tunnel): /v1/broadband/lookup returns every ISP serving an address with technology code + max download/upload speeds + low-latency flag + business/residential code; /v1/septic-score/lookup returns the v2 rural_septic_score for any TX address with component breakdown; /v1/rural-leads/county returns a ranked list of rural-septic candidates by county.

All three are backed by the 924-million-row FCC Broadband Data Collection table (116M unique US broadband-serviceable locations, 2,135 ISPs across all 50 states + DC + PR), partitioned by state for sub-second range scans. Address resolution flows: input address → property_sales fuzzy match → block_geoid → BDC range scan → JSON response. Verified production curls: FM 1696 Huntsville returns 14 ISPs and max 5000 Mbps fiber; 5160 FM 55 Italy returns rural_septic_score 85 with "Strong rural-septic indicator" interpretation.

Parcel Screen pulls parcel facts (owner, acreage, zoning, general plan) from Esri FeatureServer endpoints, runs SB-684 eligibility logic and SB-9 lot-split yield calculations, and cross-references a 776M-row CA permits database for prior approvals. Every screen produces an audit log so brokers can defend the result.

Phase 1 shipped 2026-05-13 with the core eligibility engine. Phase 2 will add statewide overlay layers (fire, flood, wetlands) and a Mapbox parcel viewer.

A Leaflet-rendered heat map and density overlay of 159K septic permits across Central Texas, centered on the Bastrop / Cedar Creek base and extending across Travis, Hays, Comal, Caldwell, Bastrop, and Williamson counties within a 120-mile fleet radius. Service-area rings (ideal <15mi, mid 15–30mi, far 30–45mi) and a dump-site inventory layer support route and fleet decisions.

The same engine can be repackaged per-operator as a "service territory intelligence" SKU at $500–1K/mo with competitor-density overlays.

A 12-route static Astro site (home, about, contact, service pages, and service-area pages) for the Honey Wagon Services brand in Rock Hill, SC, covering septic, portable toilet, and grease trap pumping. It ships zero client-side JavaScript, so it is fast by construction and costs almost nothing to host.

The brand runs on its own local phone line and its own identity, so it can run local and paid search without competing with Mac Septic listings in the same market.

A custom pixel fires on macseptic.com → POSTs to /api/public/v1/track → a 5-minute APScheduler cron identifies abandoned sessions for known returning permit owners → Brevo dispatches one of three templates (pumping_by_city, emergency, default) → an HMAC-signed vid token + 7-day suppression cap prevents over-mailing.

Alembic 126 added the schema. Sender is dannia@macseptic.com. The same first-party pipeline can be extended to permits.ecbtx.com visitors and to SMS as a second channel.

RingCentral webhooks + Twilio SMS pipe into Deepgram nova-3 for transcription, then Haiku runs forced-tool-use triage to score every interaction 0–100. If score ≥ 70, Sonnet drafts a contextual reply. DNC enforcement runs at the suppression layer so any reply path respects opt-outs.

A weekly Opus strategist email summarizes sentiment, surfaces top themes, and recommends campaign adjustments. The first Sunday Opus run was 2026-05-03 06:00 CT against the Dannia campaign `email-openers-spring-2026`. AI spend is hard-capped at $25/day with auto-pause.

Weekly `python3 -m scripts.ai_rank_check` measures position on 38 commercial keywords (e.g., "septic pumping Nashville", "emergency septic repair TN"). Execution levers are on-page optimization, a content calendar, link velocity from honeywagonservices, and AI copywriting for keyword-targeted landing pages.

The goal is top-10 on 10–15 commercial keywords in the 100-day window (2026-05-11 → 2026-08-19). A mid-window pivot is planned if no movement appears by day 60.

A 297-million-row code-violations table loaded into the same permits database. Each record carries jurisdiction, parcel, violation type, status, and date. When joined to property_sales (97M), property_liens (37M), and the permit graph (777M), it forms a national distress-signal layer.

Loaded incrementally from county and municipal portals via the gov-data-extraction framework. The dataset is the basis for a forthcoming "distress leads" product targeting field-service contractors and real-estate distress investors.

Four tables form the property side of the platform: property_sales (97M rows, 23 GB), property_liens (37M rows), acris_parties (33M NYC ACRIS records), and property_valuations (9M rows). All four are the join targets for the enrich_norm_addr.sql and enrich_with_sales.sql cron jobs that hydrate hot_leads with owner and price data each morning.

Original acquisition was a mix of BatchData purchases, county assessor scrapes, and the NYC ACRIS public bulk feed. The combined suite means any permit address in the platform can be enriched with a real owner name, last sale price, lien status, and assessed value in milliseconds.

A locally-hosted LLM cluster running on the R730 (88 cores / 768 GB RAM, dual NVIDIA GPUs, GPU-aware fan control). The active reasoning model is Qwen 3.5 122B at 81 GB on disk; nomic-embed-text handles embedding workloads for the RAG service in security-stack. Both are served by Ollama and reachable across the Tailscale mesh.

The cluster is the substrate for any non-real-time AI workload: nightly summary jobs, batch transcription scoring, embedding generation for the SOC RAG, OCR post-processing on tank inspections, and reasoning over scraped permit corpora. Anthropic Claude is reserved for latency-sensitive customer-facing flows; everything internal flows here.

Tank Vision is a self-improving OCR pipeline. It takes county septic-permit PDFs (currently Comal County CCEO in Texas), runs dual-GPU OCR, applies a learned-rules extractor that grows over time (learned_rules.json, learned_examples_comal.json), and writes structured rows directly into the permits database.

The pipeline is the cost-avoidance answer to county PIA requests like the Kendall County $3,456 quote (declined April 2026). Each new county adds incremental scrape effort but zero ongoing per-record cost.

Active coverage now spans Texas + Tennessee: Comal County (TX) plus a Tennessee cohort across Hamilton (Chattanooga), Davidson (Nashville), Rutherford (Murfreesboro), Shelby (Memphis), Williamson (Franklin), Wilson (Lebanon), Maury, and a Johnson City 4-county bundle. Hays, Williamson (TX), and Bexar are next in queue.

A standalone orchestrator on R730-2 that runs three tiers of enrichment over the permit graph.

Tier 1 is pure SQL: address normalization, exact-match joins against property_sales and contractor_licenses, PostgreSQL trigram fuzzy matching with auto-accept above 0.7 and an escalation queue between 0.4 and 0.7. A typical run enriches ~4,600 leads in under 10 minutes.

Tier 2 escalates the trigram-ambiguous matches to Claude Haiku in batches of 20; Haiku accepts above a tuned confidence of 0.85 and rejects below 0.4. Anything in between escalates again to Claude Sonnet with a 0.7 accept threshold. Every LLM call is logged with token counts and dollar cost in a SQLite ledger (costs.db) for per-day and per-month budget reporting.

Tier 3 is rich Sonnet analysis on the highest-value leads, plus a description miner (Haiku-only) that extracts septic-tank size and inspection-status hints from free-text permit descriptions. The agent also runs an anomaly detector that compares today's tier-1 stats to the 7-day median and fires a RingCentral SMS to the operator on underflow (today=0 vs median 1 already caught a regression on 2026-04-24).

The nightly sync pushes enriched rows from R730-2 to T430 with a regression gate: if any field has lower coverage on R730-2 than T430, the sync aborts and alerts. Twilio is used for phone validation on top-priority phones; a website finder enumerates contractor URLs and classifies them. The whole thing is flock-guarded so cron drift can't fire two instances.

Cron is paused since 2026-04-24 after a Tier-2 batch returned 2,250 consecutive Anthropic API errors. The Tier-1 SQL pipeline is independently working and ready to re-enable; Tier-2 needs API stability work or a fallback to the on-prem Sovereign LLM cluster (Qwen 3.5 122B) to bypass external API risk.

A full self-hosted security stack with three composed pieces: Wazuh SIEM (1514/1515), Elasticsearch (9200), and a RAG service backed by Qdrant for AI-assisted alert reasoning. A SOC dashboard fronts the alert queue and lets the operator query the RAG with natural language ("which hosts ran sudo today and have an unfamiliar process tree").

A dedicated pentest folder holds offensive tooling for periodic internal exercises. The whole stack runs in Docker Compose with separate compose files for SOC services and RAG service.

Dumpster Leads is a roll-off dumpster lead-gen product at dumpster.ecbtx.com, hosted in the permitlookup-web Next.js 16 / React 19 / shadcn app on Vercel. The customer-facing leads dashboard filters dumpster prospects by pool, E4, and permit signal off the 1.5B-record warehouse, with mobile-first responsive design and a password gate to wall off pre-launch beta access. A per-session event tracker writes to a dumpster_events Postgres table on Tailscale (event types: gate_unlock, page_view, filter_apply, lead_click, phone_tap, plus smoke-test variants) with full visitor + session + geo + viewport instrumentation. An admin analytics dashboard at /admin/dumpster-analytics renders KPIs (visitors today, sessions today, events 24h, last event), a 30-day series chart, geo distribution, the 5-stage funnel, hot-leads table, and a live event feed. The full rollup is refreshed every 5 minutes by a server-side cron (dumpster_analytics_dump.sh on R730) that runs the SQL and writes a single JSON snapshot to dashboard-cache served verbatim via Tailscale Funnel; Vercel proxies that JSON via /api/admin/dumpster-analytics. The architecture is the canonical pattern: customer-facing Next.js on Vercel, data on Tailscale-internal Postgres, analytics by static-JSON proxy — fast, secure, zero DB exposure.

Conductor AI is the internal orchestration layer for the portfolio. It spawns autonomous Claude Code worker processes in parallel, detects completion signals via promise-tag parsing on streamed JSON output, manages task queues, and runs a Rich TUI dashboard plus REST API on port 8787 with an embedded PTY terminal. Conductor monitors production sites (drphil.ecbtx.com, react.ecbtx.com, others) and triggers auto-fix sprints when health checks fail. The Gen4 experiments add a blackboard for cross-agent state sharing, a metacognition layer, adversarial verification of outputs, and a context distiller for long-running sessions. Conductor is integrated into Claude Terminal for one-command project switching. Stack: Python 3.14 + FastAPI + Uvicorn + SQLite (async) + Rich + Typer + Claude CLI subprocess workers.

This repo is the Conductor AI dashboard frontend. GitHub remote is wburns02/conductor-dashboard. The package.json is private: true, version 0.0.0, with no exports field; per the deep-doc audit there are zero downstream import consumers. Layout.tsx hardcodes Conductor branding and navigation; the "project launcher" in src/pages/TerminalPage.tsx is a menu for spawning Claude sessions across the Conductor-managed surface (DrPhil, NurseStation, ClaudeBall, PermitAPI, Conductor AI, ClaudeBallGFX), not a list of products that import this library. Highlights: live xterm.js terminal emulator with Framer Motion animations and web-links addon, advanced animation patterns (performance sprint, circuit-breaker UI), and a component framework under src/components and src/features. Stack: React 19 + TypeScript + Vite 7 + Tailwind 4 + Framer Motion 12.34 + xterm.js + React Router DOM 7. Docker-containerized via Node serve on port 3000; Playwright test coverage.

ECBTX SEO Service is an internal Docker-containerized microservice processing SEO content at scale for portfolio properties. The architecture is stateless with separate data, models, and credentials directories; nginx serves as reverse proxy / load balancer for multi-tenant or multi-endpoint routing. The service powers content automation for permit lookup, county-specific landing pages, and other SEO-heavy properties in the portfolio. Deployed on R730 with Docker Compose; version managed via version.txt; credentials isolated.

Team Edition extends HealthForge into a multi-tenant product built for a university rugby club. Coaches get training-load dashboards, alert rules, and a matchday mode; athletes log sessions and receive auto-generated training plans; an Apple Watch companion handles in-session capture. Coach views are aggregate by design so individual health data stays with the athlete.

Phases 0 through 3 (teams, session logging, coach console, alerts, matchday mode, and the watch companion) were built and shipped in a single overnight run on 2026-09-08. The watch app was confirmed on a real paired device through TestFlight on 2026-09-14, and the backend suite stands at 409+ passing tests.

Meeting Copilot listens to call audio, transcribes it in real time with Deepgram nova-3, and every 30 seconds or so sends the rolling transcript to Claude Haiku for suggested talking points, open questions, and a live tracker shown in a browser UI.

Each meeting type is a profile: a JSON file that defines goals and what to track. Profiles are re-read on every request, so adding a new kind of meeting is a file drop with no restart.

Watchful is custom hardware and firmware built on a Nordic nRF9160 cellular module. It watches aerobic-system pump pressure and the OEM alarm signal, publishes over mutually authenticated TLS to a self-hosted EMQX broker, and a bridge in the Mac Septic CRM turns failures into SMS alerts and dashboard events.

On 2026-08-26 the fleet moved from a public test broker to the self-hosted broker with certificate-scoped per-device access control. On the bench unit both critical alert paths, air-pump failure and OEM alarm, have fired end to end, including live SMS delivery to a technician's phone.

Built with a construction-industry collaborator. The concept demo maps 7 operational lanes and 10 clickable handoff scenarios where an agent tracks open items and follows up before they slip, all on simulated data. It was built and deployed the same day as the first pitch conversation, 2026-07-03. A sub-demo runs a 10-point certificate-of-insurance validation against a sample insurance exhibit.

The follow-on was a single-tenant proof-of-concept CRM, built 2026-07-15, that implements the first real workflow: chasing subcontractor compliance documents against a project's insurance requirements.

A single-page calculator that prices a pump-out by gallons and adds a distance surcharge from the real, traffic-aware Google Maps route between the company's truck yard and the job. When live routing is unavailable it falls back to ZIP-centroid distances, and then to manual miles.

Drive times are calibrated against about 25 reference routes across the 18-city service area, with metro-level handling for Austin and San Antonio. Built and shipped on 2026-09-03.

The scheduler builds agent shift schedules with an OR-Tools CP-SAT model over queue, skill, and demand constraints, then exposes them in a drag-to-edit calendar with Erlang C volume forecasting, shift swaps, time-off handling, and scenario comparison.

Tuned against a 175-agent roster it reaches 99.4% shift-floor fill, 97%+ per-queue coverage, and zero unassigned shifts within a 420-second solve budget; splitting the solve into four parallel shards cut runtime from 162s to 42s. The v2 build shipped with 44 API endpoints and 161 automated tests.

Reads two live sources per session (the session registry and each session's own transcript JSONL) for a real-time board of what's running, and separately scans every transcript on disk plus kimi CLI session logs for a day-by-day usage history, independent of which sessions are still alive. A short server-side cache keeps the expensive historical scan fast on repeat loads without going stale for more than 15 minutes.

Built and hardened the same way any production change here goes out: full existing test suite green before touching the live systemd service, a throwaway-port smoke test against real data first, then a restart and a live re-verify after deploy, never the reverse.

Runs a fixed set of tracked questions against all four major search-grounded AI providers every week, scores whether each property is mentioned/recommended/cited, and breaks the result out by market (TX/TN/SC) so a regional dip is visible instead of averaged away. Generated fresh each run, no manual data entry.