Systems from 0→1.
Owned end to end.

Departed when the company made the strategic decision to offshore its engineering team, shortly after the SCM pilot went live.

• Sole architect and tech lead of SCM: took it 0→1 to a live production pilot with a large customer while owning engineering, product strategy, and forward-deployed customer work.
• Designed and built A2UI, a declarative agent-to-UI protocol letting connected agents render live React components from a bounded LangGraph activity library and the user's capability manifest — every emission resolves through the per-invocation grant model, so agents cannot widen the FERPA surface.
• Re-architected the FERPA boundary from in-flight scrubbing to structural isolation: the model sees only opaque per-run references and numeric metrics, while student PII rehydrates in the authorized user's browser behind a per-request check. Names never reach the AI region.
• Built the capability registry, skill launcher, and per-invocation authorization model end-to-end (NestJS, Prisma, Postgres, Next.js, React): time-bound grants scope to a staff/capability/classroom/material tuple, resolve server-side, and back 7 FERPA-mapped audit events.
• Built SCM into an identity provider for connected AI agents on Python/FastAPI + LangGraph: single-use grants read scope from the database row, never the agent's request, and a per-failure-class retry policy caps each workflow at ≤2 LLM calls.
• Engineered the Temporal runtime with human-in-the-loop approval gates: deterministic, replayable workflows carry a re-validation closure, so an authorization that expires mid-review can never write to the database once approval lands hours later.

• Led the freezer-car modernization that reversed an 89% revenue-loss trajectory into $8M+/year at a 45% efficiency gain: a deterministic Kafka correlation engine joining wayside, AEI, and CAD signals on trainID, closing sensor-coverage gaps and applying an ML-derived health threshold to flag derailment-risk bearings for crew intervention.
• Designed the operator-facing consoles for Automatic Equipment Identification and Wayside Detection, replacing raw log-scraping with consistency-aware React UIs and live status surfacing — yard staff read equipment health at a glance, and data accuracy rose 85%.
• Owned the CADX web applications end to end, the interface train crews read compliance-critical directives through, from track-blockage bulletins to severe-weather authorities, where a misread screen is a safety event; also owned those features in the 2M-line legacy C++ backend.
• Fixed CADX's auth architecture after it exposed SSO session cookies to the browser: routed the frontend through a server-side proxy holding the session, moved to httpOnly cookies, and added CSRF double-submit protection so tokens never reach client JavaScript.
• Shipped field inspection and reporting tooling as an offline-first, touch-first hybrid app on ruggedized Zebra handhelds, built for gloved hands, glare, and no signal; edge caching and async sync cut inspection and reporting time 60%.
• Built a management-scoped MCP-based AI assistant over the reconciled dataset, collapsing a day-to-week analyst SQL lookup across three segregated databases into a real-time answer — field staff now locate a humped freezer car fast enough to save cargo that previously spoiled unlocated.
• Decomposed a legacy C++ monolith into modular services, a 70% gain in real-time responsiveness across three applications, and stood up structured telemetry at >90% coverage, cutting downtime 50% and deploys 40% faster.

Founded AetherisVis and sold it to a commercial gym brand within the same month, delivering the flagship product as sole full-stack engineer.

• Shipped a consumer-facing 3D product configurator for a luxury gym brand as sole full-stack engineer (live at spxgymdesign.com): buyers customize steel finishes, upholstery, and stitch patterns with instant 4K/8K feedback and smooth multi-angle zoom/pan, translated from a designer's pixel-perfect comps into production React/Next.js.
• Delivered it as a one-week commercial build embedded in SPX's sales motion, not a demo: client teams present it with no engineer in the room, configurations save and reload, and it degrades gracefully on a hotel-lobby iPad.
• Built a data-first component framework (35+ reusable React/TypeScript components) covering every configurator's shared anatomy — scene management, camera controls, material selectors — turning the SPX build into an assembly problem, not an invention one, and cutting UI prototyping time 60%.
• Engineered a frame-budget scheduler for the renderer: material swaps preview instantly at proxy resolution while full 4K/8K textures stream in only on idle frames within a fraction of the ~16ms budget, holding 60fps under concurrent camera drags and material changes.
• Compressed 8K TwinMotion renders to WebP for 70% faster loads with stitching detail intact at 150% zoom, then added lazy loading, hardware-accelerated pan/zoom, and ARIA/focus management for a touch-friendly, accessible experience.

Reached 200+ daily active users and secured a TechStars invitation within four months of launch — ultimately opted not to join, staying focused on full-time engineering work.

• Took VortexeAI 0→1 in four months, building the product surface: a low-code React/Next.js node-graph editor where non-engineers compose AI workflows by dragging, connecting, and configuring nodes, with live validation and inspector panels that keep a complex execution model legible.
• Architected the orchestration engine as a directed graph rather than letting an LLM plan its own execution: topological sort and dependency resolution schedule independent branches concurrently and isolate exactly which steps invalidate on failure, making reruns reproducible at 99.9% execution reliability atop non-deterministic model calls.
• Built a dependency-aware memory router so multi-step LLM chains receive only the state each node declared it needed, avoiding the cost and latency of stuffing full history into every call while keeping failures traceable to a single graph edge.
• Drove product-market fit through 100+ customer interviews and five pivots over seven months with a three-person team, converging on a low-code abstraction that cut onboarding time 40% and scaled to 50+ deployed automation patterns.

• Shipped two Unreal Engine 5 mobile titles with cloud auth and cross-device sync for 1,500+ players (30% faster loads, 25% better retention), earning direct recognition from the Head of UI/UX at PlayStation and reaching PlayStation Studios' promotion-support stage before visa constraints halted progression. Explore the studio →

• Designed and shipped an invite-only accountability network for founders end to end as sole engineer and designer, from product definition through live production deployment: the entire design language is one authoritative stylesheet, 77 tokens and 280 component classes across 7 routes, dark-only, with hand-vendored Radix primitives.
• Wrote the feed ranker as a product argument, not an engagement optimizer: items score on obligation and time-decayed silence, so an unanswered post rises and a settled one sinks, with engagement log-damped and hard-capped beneath every other signal — then surfaced that reasoning in the UI as per-item chips, so users can read why anything is where it is.
• Pushed every business invariant into Postgres instead of React across 10 migrations, 21 functions, 10 triggers, and 30 RLS policies: an advisory lock holding a hard 500-seat cap under concurrent signups, partial unique indexes enforcing one invite per member, and a composite primary key that makes reaction farming impossible from any client.