Engineering deep-dives
How the hard problems in my projects actually got solved: the challenges, the systems thinking behind each decision, and what I learned shipping them.
Determinism Is a Feature
VortexeAI went 0→1 in four months: 200+ daily active users, a TechStars invitation, and a low-code editor where non-technical users chain AI actions like Legos. The engineering problem was the inverse of the pitch: making a system built on non-deterministic models reliable enough (99.9%) that people trust it with real work. On graph traversal over vibes, memory routing, and wearing the CEO hat while holding the architecture.
Shipping a 3D Configurator in a Week
A luxury gym brand needed clients to see their machine, in their finishes, photorealistically, in the browser. It shipped as a commercial product in one week. What browser 3D actually costs after the first minute, why the week was fast because the framework was slow, and how commercial delivery changes the definition of done.
Your Agent Shouldn't Have a Password
Seven design iterations tried to answer "where do you put the JWT?" for an autonomous workflow that outlives the user's session. Every one leaked. The answer was to put it nowhere, and what a single-use permission slip in a database row buys you: FERPA-grade auditability, structural PII isolation, and human-in-the-loop workflows that can pause for hours.
Deterministic Facts, Probabilistic Prose
Agent trajectories carry untrusted text, and that text ends up in the prompt that explains the trajectory. Building Tracewell, I treated prompt injection as an ordinary data-flow problem: deterministic Python computes every fact, the LLM only narrates. An attacker can bend the sentences but never the numbers.
Deleting a Race Condition by Construction
The assigned bug was a one-line fix. The interesting bugs were the ones nobody asked about: a fetch race that made stale data win, a cache that defeated the feature it powered, and a 2-second latency report that wasn't a query problem. On measuring before fixing, and choosing designs where the bug can't exist.