Stack: FastAPI · Supabase · Stripe · Claude API · Railway · GitHub · GoDaddy · Brevo · Beehiiv · Google Ads · Umami Analytics · Shopify + Printful
Timeline: Core product live in 11 days. Full ecosystem — SaaS, book, e-commerce, ads, legal infrastructure — in roughly 3 months.
Infrastructure: GitHub is the code vault. Railway auto-deploys on every push — no manual deploys. All secrets in Railway Variables, never in code. Two services: invisiblearc.ai (12+ file FastAPI app) and emptyjay.net running on the same platform.
Designed a multi-phase prompt system governing a manipulation detection engine. Started with a three-layer detection model and iterated through five major architectural revisions — each time stripping engineering and trusting the model more.
Key decisions: Four-law constraint system (Law 0–3) preventing hallucination of tactic names. Tier 1 (stateless, single-call) vs Tier 2 (conversation with law injection on every user message). Taxonomy loaded with cache_control: ephemeral — ~90% cost reduction on warm reads.
The through-line: Every overengineered intake form, routing layer, and wrapping architecture got deleted. The prompt got shorter. The tool got better. Every time.
Mat wrote a 40,000-word manipulation field guide, but needed to optimize it for the prompt. Distilled to ~43,000 tokens of structured taxonomy by extracting duplicate example quotes into a centralized appendix, then dropping examples entirely once the model proved it didn't need them.
Final entry structure: tactic name · definition (mechanism + psychology) · markers (observable linguistic tells). Target: 100–180 words per entry. Validated 5/5 on untouched manipulative material, clean return on clean control email.
Same taxonomy powers the published book: Invisible Architecture: A Field Guide to Manipulation (Amazon KDP, paperback + Kindle).
Four-tier Stripe subscription model: Free · $7.95/week Standard · $25/week Max · $75/week Boost. Weekly recurring billing with webhook-driven tier management in Supabase.
Pricing iterated on margin analysis — Standard dropped from $9 to $7.95 after confirming the lower price clears the psychological $8 threshold while keeping prompt cold start margins positive.
Stripe's account review flagged the product as overlapping with regulated service categories. Corrected the internal classification inside Stripe. Account cleared and fully operational.
RLS Incident: Enabled row-level security on all five Supabase tables. Login broke immediately — silent failure, no error, just a redirect loop. Root cause: backend using anon key (routes through RLS) instead of service role key (bypasses it for trusted server). Two-line fix in database.py.
Compression Bug: Token cost climbed monotonically despite a summary-and-discard architecture designed to keep costs flat. Summaries were writing correctly but the next API call was appending summary to full history instead of replacing it. Fixed in message handler. Message 50 now costs the same as message 5.
Extended Thinking Gap: Per-message costs consistently $0.08 higher than modeled, flat across all input sizes. Pattern pointed to a fixed per-call cost. Root cause: thinking tokens billing at $25/MTok but not appearing in usage_logs. Fixed by accounting for thinking tokens separately from visible output.
Built and iterated a search campaign targeting people who know something is wrong but don't yet have language to name it. Search intent does the qualifying — ads intercept people already looking.
Restructured from a single broad ad group into four targeted clusters: General · Gaslighting (highest impression volume) · Emotional Manipulation (311 impressions, $0.37 CPC) · Silent Treatment (best CTR at 3.3%). Adjusted keyword segmentation, ad copy, and headline pinning per cluster based on search terms report data.
Budget discipline: $10/day, $0.75 max CPC, Maximize Clicks. Campaign runs impression-limited, not budget-limited — correct state at this stage. Analytics migrated from Umami to GA4 + GTM for tighter ad spend integration.
Eliminated a $400K annual MedMined contract by designing a zero-cost in-house surveillance system from scratch. When the contract terminated abruptly on April 1, 2024, the interim fallback was rulers, randomly printed lab results, and highlighters. Built a Palantir EMR solution within 2 weeks to restore proper monitoring while the permanent system was developed.
Completed VA SQL programming training to enable direct extraction from the Corporate Data Warehouse. Built complex queries to capture and filter microbiology, mycology, and viral lab data from hundreds of thousands of culture results across the Western orbit of VISN 15.
Architected a multi-layer solution combining Power BI dashboards, custom Power Apps, SharePoint lists, and Power Automate flows. Solved a two-way data interaction problem: read-only Power BI visualization with an embedded Power Apps interface allowing staff to annotate data using SharePoint as the backend.
Replaced manual Excel-based HAI tracking (CAUTI, CLABSI, VAE) with an automated data architecture using SharePoint as the database and Power BI as the visualization layer. Unit-specific dashboards auto-update daily, eliminating several hours of manual staff work per week.
Designed for dual audiences: frontline staff get real-time compliance data; executive leadership gets immediate trend visibility for rapid intervention decisions. Completed advanced Power BI training including a three-day course and one-day advanced data modeling course through VA TMS.
Replaced a manual email-based Infection Control Risk Assessment process with a LEAF workflow system — reducing administrative time by 1–2 hours weekly and achieving near 100% ICRA completion compliance.
Built a custom Power BI dashboard with complex JSON integration providing real-time visibility of active construction projects and ICRA status across the entire medical center. Led cross-functional stakeholder engagement across Facility Management, Infection Control, and Construction teams to ensure adoption. Self-directed technical learning — learned LEAF workflow development through hands-on implementation with no formal training.
Replaced an obsolete macro-dependent Excel system with a mobile-first Power Apps interface for clinical staff, backed by SharePoint data collection and Power BI dashboards with automated nightly updates.
Designed for adoption: intuitive interface reduced friction for frontline staff, improving both data accuracy and audit completion rates. Dashboard gives leadership unit-specific compliance visibility and accountability metrics in real time.
18 years of clinical practice across surgical inpatient, home health, long-term care, and federal healthcare systems. The through-line is pattern recognition under pressure, complex workflow design, and data-driven decision-making at the bedside and at the systems level.
Certified in Infection Control (CIC, 2023–2028). Yellow Belt Lean process improvement. Director of Nursing (Linden Woods Village, 2020–2021). BSN, University of Providence — 3.91 GPA, Presidential Achievement Award.
The nursing career is not the headline. It is the reason the AI work is grounded in real systems, real stakes, and real data — not toy problems.
Managed the infection prevention and control program for Kansas City VA Medical Center and surrounding Community Based Outpatient Clinics. Infection Control representative for Water Safety, Environment of Care, Construction Safety, and Infection Control committees.
Served as facility resource for data flow troubleshooting and training. Designed and owned Power BI, Power Apps, SharePoint, and Power Automate infrastructure for the infection control program. Conducted concurrent and retrospective medical record reviews for infection surveillance and regulatory compliance. Participated in mortality reviews and root cause analysis investigations.
Contact: Marilyn Lewis · (816) 861-4700
Managed caseload of up to 30 patients providing direct care and daily case management. Performed complex nursing tasks in patient home environments including wound care, IV therapy, and medication management.
Contact: Debbie Levin · (425) 525-6800
Continuum of care facility: independent, assisted, long-term care, and acute rehab. Supervised budget by controlling supply costs, labor costs, and improving census. Developed employees through annual reviews, coaching, written discipline, and corrective action. Monitored compliance through nursing documentation, resident care, and pharmacy order audits.
Contact: Cody Hughes · (816) 602-1266
Performed surveillance to identify infection trends. Created programs and projects to reduce infection rates. Audited bedside staff care in preparation for state surveys. Coordinated COVID-19 pandemic response including daily process changes based on CDC and Missouri DHSS guidelines.
Contact: Lori Gallagher · (816) 935-8948
Managed consistent caseload of 30 patients. Performed nursing interventions independently including complex wound care, medication reconciliation, and patient assessments in diverse home settings. Served as preceptor for newly hired nurses and nursing students.
Contact: Tia Hansen · (503) 215-4653
Managed 3–6 acutely ill patients during 12-hour night shifts. Performed technically complex nursing tasks. Highly proficient in EPIC electronic medical records. Preceptor for newly hired nurses and BSN/ADN student nurses.
Contact: Legacy Emanuel Hospital · (503) 413-2200