Cofounded with Bradley Kipp in 2013. Feminella described directing engineering, hiring, technical strategy, and architecture while remaining hands-on; his formal title was cofounder rather than CTO. The subscription service connected to existing business analytics streams, detected changes, notified customers, suggested actions such as checking referral sources, and offered access to an analytics counselor. This is a historical project account, not confirmation of continued operation.
Feminella's portfolio attributes this 2013–2014 smart-home sensor deployment platform to his collaboration with Filippo Valsorda under Kamin Whitehouse's supervision. It supported remote deployment, self-diagnosis and recovery, updating, and data reporting to reduce technician visits to disrupted household installations. The portfolio explicitly links this repository; the retained repository page establishes the code artifact but does not independently establish individual authorship or ownership of a personal GitHub account.
Feminella states that he coauthored a paper about the smart-home deployment work appearing in the Proceedings of the 1st ACM Conference on Embedded Systems for Energy-Efficient Buildings, BuildSys 2014. His account also reports a Best Presentation award. The retained evidence supplies neither the paper's exact title nor a publication-specific URL or complete author list; this link is the first-person portfolio account, not the paper itself.
Feminella credits his team with building the system during 2007–2009, not himself as sole author. They replaced reporting through physically shipped CDs with a large Java web application incorporating hundreds of data-integrity checks. The system consolidated emissions submissions from US jurisdictions and industries into the National Emissions Inventory for researchers and policymakers. Historical statements about its continued use are not treated as present-day verification.
As a consultant in 2006, Feminella built an Asterisk VoIP server for exploratory testing of the university's next-generation telephone system. His portfolio reports capacity for routing and connecting up to 6,000 simultaneous SIP calls. This establishes his prototype contribution, not authorship of Asterisk or sole responsibility for the university's subsequent telephone migration.
During 2006–2007 consulting work for Northrop Grumman, Feminella optimized an existing navigation codebase, audited technical debt, and implemented Vincenty's method for distances on an ellipsoid to improve waypoint accuracy. This is a contribution to an existing product, not sole product authorship or authorship of Vincenty's algorithm.
As Cardagin's first CTO in 2011–2012, Feminella led a technical turnaround after a difficult launch. His account describes growing the technical team from zero to eight people in twelve weeks and collectively building a website, two mobile applications, and a business administration platform. He subsequently stepped down after the principal technical rebuilding work. This is engineering leadership and team delivery, not sole implementation.
For a 2006–2007 consulting engagement, Feminella built a Python web application to collect and analyze focus-group data for revised Girl Scout Cookie recipes. It also supported projections of recipe costs and production capacity. His contribution was the analytical software, not authorship of the recipes or proof that his application caused subsequent sales results.
jxf.me
Bio, Work & Ideas
John Feminella
Conference affiliation: Two Sigma
On this page
John Feminella is a software engineer, technology leader, and cofounder of the analytics-monitoring startup UpHex. His work spans environmental data systems, distributed sensors, enterprise cloud platforms, and applied AI. Alongside building those systems, he has helped organizations change how they develop, deploy, and operate software.
From consulting to engineering leadership
Feminella’s early consulting work involved systems with concrete physical constraints. In 2006, he built an Asterisk-based telephony server for the University of Virginia’s exploration of a new phone system. For Northrop Grumman, he improved navigation software by implementing Vincenty’s method for calculating distances on an ellipsoid—a more precise model of the Earth than a sphere. These projects combined infrastructure development with the practical demands of routing calls and plotting ships’ courses. (Selected consulting projects)
Between 2007 and 2009, he helped build the Environmental Protection Agency’s Emissions Inventory System. His team replaced a process that circulated sensitive emissions data on physical CDs with a Java web application containing hundreds of data-integrity checks. The system collected pollution data from jurisdictions and industries across the United States for the National Emissions Inventory, giving researchers and policymakers a consolidated dataset to work with. (Project account)
As Cardagin Networks’ first chief technology officer in 2011–2012, Feminella took responsibility for a turnaround after a difficult initial launch. He grew the technical team from zero to eight people in twelve weeks and led the development of a website, two mobile applications, and an administrative platform for businesses running loyalty programs. He subsequently stepped down after the major technical rebuilding work was complete. (Cardagin portfolio entry)
Turning data into useful action
Two projects from the following years addressed different reasons that collecting data can fail to produce useful results:
UpHex: notice changes without watching dashboards all day. Feminella cofounded UpHex with Bradley Kipp in 2013 and directed its engineering, hiring, and architecture. The subscription service connected to business data streams such as Google Analytics and Shopify, analyzed changes, and notified customers when something deserved attention. An unexpected traffic spike, for example, could trigger a suggestion to inspect recent referral sources. Customers could also request help from an analytics counselor. The product addressed a shortage of time and expertise to interpret data businesses were already collecting. (UpHex portfolio entry)
Piloteur: keep sensor networks collecting data unattended. Working with Filippo Valsorda under Kamin Whitehouse’s supervision, Feminella built Piloteur, a deployment platform for smart-home sensor networks. Household research installations faced power outages, pets, and other disruptions that could otherwise require a technician’s visit. Piloteur supported remote deployment, updating, recovery, and data reporting, reducing the need for researchers to service each installation in person. (Collaborative project account)
Enterprise platforms and applied AI
Feminella joined Pivotal on August 1, 2015, as an advisory field engineer while remaining a principal in UpHex. Pivotal’s commitment to open-source software appealed to him, as did a role that combined business strategy with practical engineering. Working with Pivotal Cloud Foundry, he helped enterprises adapt their software-development and deployment practices to cloud environments. (Pivotal advisory work)
His subsequent career included technology leadership at VMware and ThoughtWorks and enterprise software work at Nuvalence, where he addressed both software delivery and the organization of engineering teams. He also helped design FedNow, an instant-payments network. His professional history includes leadership at Two Sigma. (Career history and FedNow contribution)
By March 2024, his work at Nuvalence included exploring generative AI for production observability—the use of operational signals to understand running software. He later helped design and scale an AI platform at EY. (Observability work; EY career account)
On September 18, 2025, Hallmark Health Care Solutions appointed Feminella senior vice president of data and artificial intelligence. His remit was to develop data and AI strategy across the company’s healthcare workforce platforms. He described the immediate task concretely: structure workforce and compensation data for analysis, connect it to predictive services, and embed those tools in clients’ daily operations. That was an agenda for the role, rather than a claim that those capabilities had already shipped.
Deployment beyond packaging
Feminella’s Pivotal-era exploration of unikernels illustrates his approach to emerging infrastructure. Unikernels package an application with only the operating-system components it needs into a specialized machine image. He experimented with MirageOS, building a small stock-quote service, while distinguishing that exploration from academic research or production readiness. Debugging, tracing, deployment tooling, and suitable workloads remained open questions; he argued that those gaps warranted investigation rather than either immediate adoption or dismissal. (Unikernel experiments and tradeoffs)
His larger concern was the work left between packaging an application and making it available to users. Giving a developer a container still leaves orchestration and other operational responsibilities. He advocated platforms that carry an application through to a running service, so developers can return to writing software instead of managing the machinery around it. (His argument for application platforms)
John Feminella uses Citibank’s mistaken Revlon payment to explain five composable controls for autonomous agents: make them visible, review them by risk, enforce purpose-based access, share evaluation infrastructure, and reserve human attention for consequential decisions.
A warning and multiple approvers can still miss the consequential distinction: Citi intended an interest payment, while the software sent principal too. Controls must help distinguish intended actions from dangerous ones.
Make registration a low-friction part of building agents. Ownership, declared tools and endpoints support both governance and reuse; a heavy universal gate can push work out of sight.
Review agents by risk and enforce access by purpose. In managed environments, registry policies become restrictions on endpoints, hosts and resources rather than instructions the model is expected to remember.
Share the common evaluation machinery while leaving teams responsible for agent-specific expectations. A model change is a dependency change whose effects need evaluation.
Use deterministic rules for settled cases, independent classifiers for remaining routine decisions, human review for consequential ambiguity and outright denial for clearly dangerous calls.
▶ Watch ↗