Portfolio & Exhibits

The field partner for working patrol: scripts, DUI/SFST walkthroughs, cited state law, incident narratives, and a search-warrant builder that reads an officer's voice into the blanks of that county's own court PDF — then the officer reviews, attests, and swears it. Live on the App Store and Google Play at $7.99/month. Built and shipped in fifteen days: 699 commits, ~70,000 lines of AI-written TypeScript over a 275,000-line law corpus pulled from official state sources rather than aggregators, 57 cloud functions, 118 test files, and Apple, Google, and Stripe billing reconciled nightly against each store's server API.

As of September 2026: 100+ accounts, 30 paying subscriptions, 200+ followers on its Facebook page — officers on real shifts, two weeks after the first binary shipped.

"The K9 Handler's Unfair Advantage." Voice-first capture (Deepgram + Grok) generates court-admissible narratives aligned with Florida v. Harris admissibility expectations; case-law verification, offline-first SQLite mirror with Firestore outbox and per-collection watermarks. CMEK on Firestore from project creation, App Check with real attestation providers. Live on the App Store and Google Play, carried on shift by a sworn handler in Monroe County, AL. 1,918 commits and roughly 300,000 lines of AI-written TypeScript across mobile, web, functions, and a shared domain spine; 216 test files.

Resilience training for first responders: self-guided lessons, an Aid Bag of one-tap tools for hard moments, a private journal, podcasts, and video, built to the curriculum of Jim Martel, LCPC. One Expo codebase for iOS, Android, and web, a content studio so the library grows without an engineer, Stripe subscriptions and department seat billing, and a grant-reporting and impact dashboard. Client-owned from the first commit; contracted through bigBespoke LLC on a fixed-price SOW and billed by milestone against a requirements-verification matrix. 459 commits, 33 cloud functions, and 43 test files, all AI-written under my direction. Live on the App Store and Google Play.

Multi-agent LLM training platform for police de-escalation. Three-agent pipeline (Dialog → World → Psyche), automated ICAT-aligned grading (6 categories / 10 points), 10-dimension session-insight scoring, skill-tree progression with freshness decay, Bronze → Diamond certification tiers. The platform is complete and live at copapp.ai with a public no-signup demo, founder console, Stripe billing, and a 29,000-record outreach CRM built on the same functions; department rollout is paused on go-to-market resources, not on engineering. 291 commits, 95 test files, CJIS-prep posture from the rules layer up.

Atlas is an open-source LLM inference engine written from scratch in Rust and CUDA and tuned for NVIDIA DGX Spark: zero-trust inference on hardware you own. I started building at Atlas in August 2026. My first bylined post, Atlas is the inference engine from desktop to hyperscaler (September 11, 2026), argues that inference is an infrastructure problem, where more tokens per watt at a sane total cost of ownership decides who gets through the capacity bottlenecks.

1.9 M-token RAG over 384 sermons spanning 8 years of teaching, powered by Grok Collections. GraphSmart interactive timeline with drag-to-select; multi-window time-range filters (All / 3yr / 2yr / 1yr / 6mo); smart token-budget management against the 1.9 M-token ceiling; auto-detected sermon series and seasonal grouping. Built as ChatWithRob; live today as Ask Pastor Rob. A later cost pass cut the running bill roughly 87% — about $16/month down to under $2 — and the repo carries its own SEO operations and owner’s guide so the church can run it without me.

A spec-driven pitch system: one vertical spec file — palette, promise, modules, stats, AI hooks, local targets — compiles into a themed, AES-256-GCM access-gated interactive concept demo and a print-ready one-pager PDF for every target business in that vertical, plus a shared clickable prototype app that deep-links per shop. Eight verticals live — med spa, dental, eye care, fitness, home services, auto care, dog care, coffee — 48 demos and 48 one-pagers deployed on this domain. Adding an industry means researching it and writing one spec; no template edits.

A self-hosted e-signature tool, built because paying per envelope to send my own statements of work made no sense. Composer with drag-placed text, date, and checkbox fields; capability-link signer URLs; ESIGN/UETA consent captured on the record; an executed PDF plus a signature certificate carrying the audit trail; unsigned-copy download for signers; live status dashboard on recents. The Firestore rules are built around unguessable capability links rather than accounts, so a signer never has to register for anything.

The ballot you will actually be handed, every contest sourced and dated, and the join from ordinary life to the office that holds it — noise after ten, a flood, an uncollected bin each already belong to somebody's desk. Starting with Alabama, because local coverage is the whole system. Deliberately not a habit machine: no streaks, no quizzes-as-facts, no turnout training. Every published sentence carries a source and a last_verified date, and an office with no certified names yet renders as a blank — the blank is the record.

The civilian derivative of PawTrek's substrate — not a walk tracker, and not the household demo everyone builds first. It copies PawTrek's operating laws without touching PawTrek: the working pair is the subject, relationships are the affordance, events happen in the world and addenda grow the file, and the coach proposes while the handler commits. Lenses reshape the whole tab bar per role — a guide cannot invent events on a pair he does not handle, a watcher has no Ask at all — and logging a negative truth is exactly as easy as logging a win.

AI encounter rehearsal for clinicians. A living room, a clock, and a score that only clears when the person in front of you can actually leave — the end of the visit is where the life continues or doesn't. Deliberately not a CopApp reskin and not an empathy module: the simulated patient's state, the time pressure, and the grading are all built for the close of an encounter, which is the part clinical training rushes. Training and education, never a medical device.

LLM-driven middleware for federal legacy-system integration friction. Discovers heterogeneous I/O (REST / CSV / PDF / SQL / screen-scrape), generates a Python connector, auto-synthesizes pytest cases, iteratively refines until tests converge. Pluggable provider router auto-selects between Grok 4.3 (cloud) and Qwen3 via Ollama (local / air-gapped) by data sensitivity — no manual switching. AST-policy + subprocess sandboxing on every generated connector.

A physicist's take on celestial mysticism: daily readings framed in real astrophysics — orbital mechanics, constellations, stellar physics — using validated ephemerides via @tubular/astronomy. The point is the user journey: someone curious about astrology gets dropped into actual physics on the way, scaffolded so they can learn instead of being thrown at complexity.

14-translation probabilistic consensus engine over the Bible. Dynamic-programming word alignment across versions (BSB, KJV, ASV, JPS, Vulgate, Young's Literal, etc.) produces a word-by-word confidence score; the UI surfaces the consensus reading and color-codes each word from dark green (90%+ agreement) to red (< 30%) so a reader can see where translation disagreement lives. Hover reveals every variant and which versions used it.

Interactive fire simulation where fire is chemistry, not a particle effect. Every cell of the domain carries a full atmospheric state — velocity, temperature, pressure, O₂, fuel vapor, soot, steam, CO₂, N₂ — and solids dry, pyrolyze, flame, char, smolder, and ash on their own clock. Flames need oxygen and suffocate without it; CO₂ sinks and steam rises; a sealed room builds pressure by PV = nRT and vents with a jet when you open the hatch. All physics in WGSL compute, verified at 123–146 fps across 2.5 million cells, with five combustion constants tunable live while it runs.

A continuous, physics-authoritative flight surface from ground level to orbit: Earth-scale terrain and atmosphere, a real sun at real distance, real orbital mechanics, and a water-propellant SSTO you board on the ground, start up, and fly out through weather and reclamation altitudes with no loading screen and no magic physics. Multi-scale coordinate handling is the whole engineering problem — float precision that has to survive both a bolt and an astronomical unit. Hard target of 120+ fps at 7680×1440. The sibling repo flight-bench runs the pilot-in-the-loop rig at 240 Hz.

Real-time multi-agent mission-control simulator orchestrating 8 LLM-driven console operators under a Flight Director command loop. G-FOLD lossless-convex landing guidance, CHUG / L1 / T1 combustion-instability modes, JONSWAP droneship dynamics, emergent IMU / GPS / RCS / radar-altimeter failure cascades. Validated through an 8-flight Monte Carlo campaign (448 LLM calls, A− grade).

End-to-end classical GNC stack plus a self-authored 44-chapter technical book. Reference simulator: 9-state EKF at 1 kHz, PN guidance at 100 Hz, plant-inversion control on RK4 1 ms physics, tactical-grade MEMS IMU model. The book and the code were written for the same audience: a guided-system engineer who needs the doctrine and the architecture in one place.

Python rocket simulator with real orbital mechanics, neuromorphic adaptive GNC, powered explicit guidance, multi-atmospheric models (US Standard, Jacchia, custom), and Monte Carlo uncertainty quantification. Procedural rocket design across a high-dimensional configuration space with quantum-inspired optimization for engine / mass-fraction / turbopump-speed sizing. The point: a working simulator I could iterate on long-form GNC ideas inside of.

The whole planet as a true-proportion interactive disk, inner core to magnetosphere, with every major Earth system running live and coupled: a real (coarse) Boussinesq mantle-convection solver drives force-balanced plate tectonics through Wilson cycles; new seafloor records the stochastic geodynamo's reversals as readable magnetic stripes; a Budyko–Sellers climate breathes through Milankovitch ice ages; a working petroleum system births and destroys oil fields; all 118 elements live where Earth actually keeps them. Interior is the actual PREM model — verified by a 29-check acceptance suite (mass, gravity, moment of inertia, core pressures). Zero dependencies, pure canvas + vanilla JS, deterministic per seed.

A real-scale 3D simulation of the Sun, Earth, Moon, and 25,000+ stars across ±5 million years. Walk the planet's surface at any epoch and watch the Milankovitch cycles — obliquity, precession, eccentricity — reshape sunlight, trace analemmas, and see the pole star drift from Thuban to Polaris to Vega. Laskar 2004 orbital solution, Vondrák–Capitaine–Wallace 2011 precession, HYG star catalog — all original TypeScript astronomy. Zero dependencies, runs offline in the browser.

Empirical atmospheric tomography: treat every storm in the radar archive as a probe, watch how the land reshapes it, and collapse the archive into a land operator field — how each patch of ground modifies weather, regime by regime. Not forecasting; forensic decomposition of real observations, or "how the storm sees the land." The pipeline runs end to end on real archive data and has produced its first measurement.

Own a base, hangar a plane, climb, hold a banked orbit, man the guns, land, improve the base, scramble again. The engineering interest here is the honest constraint: you cannot solder four proprietary engines into one process, so every desired "touch" is mapped to a subsystem that is either implemented or legally approximated — a documented feel-map instead of a Frankenstein. Two hosts share one campaign state: a scripted high-fidelity host for urban density and ragdoll, and an owned simulation process for base meta, logistics, fuel, and parts. 248 commits, including a real orbit autopilot with a clean deorbit.

Submarine command simulation where you walk the boat rather than read a checklist. Coupled physical systems, a compromised comms lattice, and control laws that change behavior with the ocean regime you are in. You do not win by executing procedure — you win by discovering how subsystems couple under a shifting regime, then reshaping those couplings so positive emergence compounds while negative emergence stays contained. The same cascade saves or kills depending on state; nothing has innate moral valence. Every patrol cold-opens mid-scram.

An air-traffic-control sim where every pilot on the frequency is an LLM. As LAX Approach you vector arrivals with real phraseology — typed or spoken — while AI pilots with seeded personalities read back, push back, and occasionally declare emergencies that play out through freeform conversation. xAI Grok drives the interesting moments; templates carry routine calls; voices run through a VHF radio-filter chain. The radar scope is a four-layer Canvas renderer with phosphor decay and CRT post-processing. A second Steam edition runs the whole frequency offline — whisper-class ears, piper-class voices, and a small first-officer model distilled from the connected edition’s pilots on synthetic data. No account, no key, no internet after install.

PyVista / PyQt6 desktop tool that converts an arbitrary STL into a lattice geometry: cubic, gyroid TPMS, auxetic-honeycomb, or auxetic-star unit cells, with curvature transformations and continuous density gradients. Backed by Johnson-Cook plasticity / Hopkinson-bar material requirements for additively manufactured metals — so the lattice geometry you generate is one your printer and your material can actually produce.

STL / STEP / IGES analysis with first-principles hydrodynamic calculations (hull speed, drag coefficients, Froude numbers) and printability / machinability / moldability scoring backed by LLM-powered design recommendations. The geometry tier is real; the LLM tier sits on top to translate "this design will warp at corner X" into "consider chamfering corner X to radius Y."

Turns a personal media library into a broadcast television station: numbered channels, continuous schedules, and the lost pleasure of arriving — flipping to a channel and landing twenty minutes into something you love, no menu, no decision, already playing. Every channel has been airing since the epoch and nothing is ever paused, because a broadcast does not wait for you. Tune in and it computes exactly what its schedule says is on right now, then drops you there mid-stream.

Share one mouse and keyboard across Windows machines over the network: move the cursor off the edge of one screen and it appears on the other, keyboard and clipboard following. One Python file, no dependencies, no cloud, no account — written after the third commercial option wanted a subscription for something that is fundamentally a socket and an input hook.

Every statement you download becomes one searchable, categorized ledger and a dashboard anyone on the same wifi can open in a browser — built so a parent could see the whole picture without learning a product first. Local only: no bank credentials, no aggregator, nothing leaves the machine. The categorization is the hard part, and it is deliberately correctable by hand, because a ledger you cannot correct is a ledger you stop trusting.

Headless agent control of Adobe Premiere Pro for long-form rough cuts: import, conform, bin, build a sequence, scatter clips with breathing room, then hand a real project to a human who tunes it in the live GUI. The interesting work is archaeological — the application's sealed DOM and QE layers taken apart into a grammar and a set of joints an agent can address, with a second will breathed into the running app rather than a script fired at a closed file.

Multi-source weather-forecast trading system fusing GFS ensembles, NWS METAR observations, and the unstructured-text NWS Area Forecast Discussion via a multi-agent Grok pipeline. Kelly-criterion sizing with a paper-trade validation gate before any live capital. Built to find out what an LLM is genuinely good at when given fundamentally heterogeneous evidence — and where deterministic numerical forecasts still beat language models cleanly.

Private real-time WebRTC voice platform with an always-on Grok voice agent, "Eve," using xAI Collections for cross-session memory. Built to learn what an always-listening voice agent feels like when it actually remembers you between calls — and what the latency / privacy tradeoffs look like at the architecture level.

A data platform that reduces every U.S. political figure and issue to a 7-dimensional octonionic vector: 3,077 figures, 10,300+ issues across 250 years, $16.8B in documented campaign finance, and ~13,000 mapped connections. Swipe issues to build your own vector (Kalman-filtered) and see who you align with by math rather than party. ~12,770 statically-generated pages live at prism.expert.

Extended abstract for AI4SE 2026, continuing the INCOSE line of work on quantifying communication in high-reliability organizations. The argument: the people-organization layer is treated as context rather than as a modeled subsystem, and agentic tooling now makes it tractable to infer that layer's structure from how people actually interact — schema from natural interaction, no questionnaire. CopApp serves as the worked example: the same agents that grade de-escalation infer a contributor's thinking-talent profile from how they respond under pressure.

Unverified rev 1 of a 13-paper, 241-page working collection exploring whether the Yang–Mills existence and mass gap problem can be approached through a non-associative algebraic framework grounded in imaginary octonions and Malcev algebras. Opens with a Jacobi-Alternativity No-Go Theorem, builds through contextual PBW, decompactified Killing form, Moufang–Kolmogorov spectral theory, tree-filtered Sobolev estimates, lattice gauge-scalar reflection positivity, and a Feshbach–Schur mass gap argument. Complement to — not contradiction of — the Wheel-Laurent work that refutes prior overclaims. Working draft, not yet circulated for peer review.

Interactive 3D visualization exploring sine waves as projections of helical motion in unbounded solution space. Five resolution theorems showing how classical wave oscillations resolve into uniform circular motion with constant energy when the truncated dimension is restored. Three.js scene with real-time particle systems, energy-flow visualization, and multiple camera views.

Peer-review-grade synthesis of Carlström's wheel-of-fractions with formal Laurent series. Contributes a total symbolic algebra distinguishing four types of zero with provenance carried through indeterminate forms — and explicitly refutes prior overclaims on Yang-Mills resolution. Python computational verification suite covering Cayley-Dickson, sedenions, Puiseux, and Borel summation.

A formalization program working the Cayley-Dickson ladder past the octonions and asking what survives each doubling — and what the loss of associativity, then alternativity, then composition actually costs a physical theory built on top. Written as a research program with a hundred candidate propositions and an explicit falsification path rather than a manifesto: the value is in which candidates die, and why.

Open-source experimental tool for AI alignment auditing via octonionic-signature representation-shift detection. The premise: in higher-order algebraic embeddings, a model's representation of "the same concept" before and after fine-tuning leaves a measurable signature; the divergence of that signature is a candidate misalignment indicator. A research probe, not a product — but the kind of thing that becomes a paper.