HTML version — the complete career record, no JavaScript required

Geronimo Moralez IV

Staff Machine Learning Engineer AI Systems Architect Technical Director

Builds infrastructure for creative people.

→ Interactive version (canvas, keyboard-navigable, honours prefers-reduced-motion)

Executive summary

Staff ML engineer and systems architect with 18+ years spanning machine learning, VFX, XR, real-time graphics, and voice synthesis. Designs and operates production AI systems end to end — sovereign GPU inference on in-house Blackwell hardware, multi-agent orchestration, LLM safety engineering, and cloud-native product delivery — including a live consumer AI product and paid enterprise modernization work. Known for adversarial self-audit discipline: systems ship with measured claims and verified outcomes, not vibes.

Geronimo Moralez IV is a staff machine learning engineer and systems architect in Los Angeles. His 18+ years span machine learning, VFX, XR, real-time graphics, and voice synthesis — a range that started in stereo conversion and pipeline engineering and now runs through sovereign GPU inference and multi-agent production systems.

The through-line is infrastructure for creative people. At StereoD he helped build the stereo-from-depth workflows the company scaled on. At Sony Pictures Imageworks he integrated the studio's first Alembic geometry pipeline, working with the format's original authors. At Encore Hollywood he automated compositing and render-farm work across thousands of episodic shots. The tools changed; the job did not.

Nimo Labs has run alongside all of it since 2017 — a boutique R&D studio delivering AI systems, voice synthesis, XR applications, and cloud-native platforms for clients, alongside an internal portfolio in agentic engineering, applied ML, and consumer AI. The inference hardware is in-house: four NVIDIA B200 Blackwell GPUs and an RTX A6000.

He is known for adversarial self-audit discipline. Systems ship with measured claims and verified outcomes, not vibes — which is why every number on this site carries a provenance tag, and why the figures that could not be sourced were deleted rather than styled.

Signature strengths

MULTI-AGENT AI SYSTEMS
Orchestration runtimes, coordination infrastructure, verified production pipelines.
SOVEREIGN LLM INFRASTRUCTURE
Trillion-parameter-class MoE serving, INT4 quantization, vLLM, unit-economics analysis.
SEMANTIC SEARCH & KNOWLEDGE GRAPHS
Graph-RAG across Neo4j, Qdrant, and multimodal embeddings.
LLM RELIABILITY & SAFETY ENGINEERING
Error-corrected execution, verification schemes, statistical certification, crisis-detection pipelines.
VOICE SYNTHESIS & AUDIO INTELLIGENCE
TTS optimization, prosody, neural vocoders, and audio intelligence pipelines.
REAL-TIME GRAPHICS & SPATIAL COMPUTING
Physics simulation, shader development, and XR across Unity and Unreal.
PROTOTYPE TO PRODUCTION
Rapid prototyping through to production in high-ambiguity environments.

Professional experience

Crafty Apes VFX

VFX Pipeline & Machine Learning Engineer

– Present · Remote

Semantic asset-discovery for VFX workflows — retrieve by meaning, motion, texture, composition, and visual energy.

  • Designed semantic asset-discovery systems for VFX workflows using multimodal embeddings, visual feature extraction, metadata enrichment, and vector similarity search.
  • Developed artist-aligned taxonomies and ontology layers translating concepts such as motion, texture, composition, and visual energy into machine-readable retrieval signals.
  • Built indexing, ranking, and evaluation pipelines enabling fast, meaning-based search across large visual-asset collections.
  • Prototyped AI-assisted workflow integrations connecting machine-learning capabilities with established VFX tools and production pipelines.
  • Collaborated with creative and technical stakeholders to translate ambiguous production needs into testable ML systems and practical artist-facing tools.

Impact & outcomes

  • Bringing ML superpowers to Artists, Production, and Technology.

Stack Multimodal embeddings Visual feature extraction Metadata enrichment Vector similarity search Artist taxonomies Evaluation pipelines

  • Oct 2024 – Present.

Nimo Labs LLC

Founder & Principal Technologist

– Present · Los Angeles, CA Persistent lane

The persistent lane. A boutique R&D studio running underneath every other engagement since 2017 — client delivery on one side, an internal portfolio in agentic engineering, applied ML, and consumer AI on the other.

  • Boutique R&D studio delivering AI systems, voice synthesis, XR applications, and cloud-native platforms for enterprise and startup clients — alongside a proprietary internal R&D portfolio in agentic engineering systems, applied ML, and consumer AI.

SOVEREIGN LLM INFRASTRUCTURE

Trillion-parameter-class mixture-of-experts inference, served from hardware in the building rather than rented by the token.

  • Deploy and operate trillion-parameter-class MoE inference (INT4/AWQ quantization, vLLM) on in-house 4× NVIDIA B200 Blackwell and RTX A6000 hardware.
  • Benchmark-driven serving optimization with measured 1,100+ tok/s sustained throughput and self-hosted unit-economics analysis.

MULTI-AGENT PRODUCTION SYSTEMS

Agentic engineering infrastructure operated end to end, not demonstrated — the orchestration layer these products ship on.

  • Design, build, and operate agentic engineering infrastructure end to end — orchestration runtimes, model-routing gateways, coordination messaging, automated verification with evidence-anchored audit trails, and human-in-the-loop approval gates.
  • Used daily to ship real products.

LIVE CONSUMER AI PRODUCT

An LLM-powered companion app in production. The product stays unnamed here, matching the résumé's disclosure level.

  • Built and operate an LLM-powered companion app in production on AWS serverless — 17 CDK stacks across Cognito, DynamoDB, CloudFront, WAF, and SES.
  • Defense-in-depth safety stack: fail-closed age gating, multilingual crisis detection with automated production canaries, jailbreak firewall, moderation and abuse ladders.
  • Flutter client with Cognito SRP auth, Google SSO (PKCE), and Google Play Billing with server-authoritative entitlements.

APPLIED ML RESEARCH

Graph neural networks pointed at creative intelligence, and the corpus and reliability work that has to exist underneath it.

  • Graph neural networks for creative-intelligence applications (PyTorch Geometric, CLIP joint embedding spaces, perceptual color science).
  • Large-scale training-corpus engineering with adversarial leakage auditing.
  • LLM reliability research — error-corrected long-horizon execution with statistically certified reliability bounds.

CLIENT DELIVERY — LEGACY MODERNIZATION

A twenty-year-old WebForms platform replaced with serverless, on a fixed fee, in production, with PHI in scope.

  • Replaced an ASP.NET 3.5 WebForms (VB.NET) + SQL Server platform with a React/TypeScript SPA on AWS serverless — 37 Lambdas, 33 routes — for a San Diego legal-services firm (2026).
  • Fixed-fee milestone engagement, delivered and paid in full, carrying a HIPAA/PHI workstream.
  • Earned a scoped autonomous-deploy authorization in production.

VOICE / TTS CONSULTING

Production voice synthesis for clients — prosody, vocoders, and the inference cost of both.

  • Production voice synthesis systems — prosody optimization, neural vocoder integration, iOS native speech framework implementations ($25K+ engagements).
  • Measured 30–50% TTS inference speedups via FP16 and KV-cache optimization.

A.R.I.E.S.

A multi-agent research system that reads the literature so the operator can interrogate the synthesis instead of the corpus.

  • Multi-agent autonomous research system performing literature review, synthesis, and knowledge extraction.

Stack vLLM INT4 / AWQ NVIDIA B200 Blackwell RTX A6000 PyTorch Geometric CLIP Neo4j Qdrant AWS CDK Cognito DynamoDB CloudFront WAF SES Flutter React / TypeScript

  • Specifics available in conversation, under NDA where required.

Immuse

Tech Lead — AR Music Platform & Web3 Systems

· Remote

Engineering lead on an AR NFT music platform where spatial computing, on-chain asset management, and live performance capture met — a role spanning AR, blockchain, cloud architecture, machine learning, and project management at once.

  • Led engineering for an AR music platform integrating spatial computing, blockchain infrastructure, and real-time performance capture.
  • Developed NFT minting pipelines, smart-contract tooling, and Web3 asset-management infrastructure.
  • Built native iOS plugins for AR positioning, spatial audio, and immersive fan-engagement experiences.
  • Directed cross-functional teams spanning product, creative, and engineering through multiple release cycles.
  • Led a team of four engineers across web3, backend and Unity, running very active agile sprints.
  • Visual music-based AR NFT development, on an augmented reality pipeline built in C# and Python.
  • Audio neural-net integration with ONNX and Unity Barracuda, and a Visual Positioning System implementation.

Impact & outcomes

  • Shipped AR location-based music NFTs — music anchored to physical places, minted on-chain.

AUGMENTED REALITY PIPELINE

Visual music-based AR NFT development — the spatial layer the product performed in.

  • Visual music-based AR NFT development.
  • Augmented reality pipeline in C# and Python.
  • Visual Positioning System implementation.

AUDIO NEURAL NET

Neural audio inference running inside the Unity client rather than behind a service call.

  • Audio neural-net integration with ONNX and Unity Barracuda.

ON-CHAIN SYSTEMS

Contract deployment and live operations, plus the wallet, storage and node infrastructure around them.

  • Solidity contract deployment and live operations.
  • ThirdWeb, IPFS, Consensys NFT, Infura, MetaMask, Alchemy, and Moralis.

BACKEND & DEVOPS

AWS serverless, deployed through CloudFormation.

  • Backend DevOps and development — AWS serverless with CloudFormation deployments.

NATIVE iOS BRIDGES

Swift plugin bridges between the engine and the platform.

  • Native iOS Swift plugin bridges.

Stack AR / spatial computing Unity (C#) Python ONNX / Unity Barracuda Visual Positioning System Solidity AWS serverless CloudFormation ThirdWeb IPFS Infura Alchemy Moralis MetaMask iOS native plugins (Swift) Spatial audio

  • May 2022 – Oct 2023.
  • A complex position spanning several technical domains at once: AR, blockchain, cloud architecture, machine learning, and project management.

Chicken Bone FX

Pipeline TD — Contract

· Los Angeles, CA

A short pipeline contract: Nuke and ShotGrid, plus scalability and telemetry consulting for deployment and monitoring.

  • Nuke and ShotGrid pipeline.
  • Scalability and telemetry consulting for better pipeline deployment and monitoring.

Impact & outcomes

  • Stood up the studio's editorial pipeline end to end.

Stack Nuke ShotGrid Telemetry

  • Nov 2021 – Apr 2022.

Apple SPG

Real-Time Pipeline Developer — Special Projects Group

· Cupertino, CA

Real-time rendering and AR material R&D inside Special Projects, at résumé-level disclosure and no further.

  • R&D on next-generation AR material systems, shader optimization, and real-time rendering pipelines under strict confidentiality.
  • Developed physics-based water simulations for Apple Watch interfaces — shipped in production firmware.
  • Collaborated with research and product teams on advanced spatial-computing workflows and visual-fidelity benchmarks.

Impact & outcomes

  • Architected a unified shader pipeline.

Stack Shader optimization Real-time rendering Physics simulation

  • Mar 2021 – Oct 2021.
  • Résumé-level text plus one operator-dictated impact line (2026-08-10). The confidentiality line on the résumé remains the ceiling for everything else.

Digital Giant

Tech Lead — Automotive Real-Time Pipeline

Established an automotive pipeline with cloud pixel streaming and batch processing against manufacturer APIs, leading a combined art and engineering team.

  • An automotive company where I established an automotive pipeline with cloud pixel streaming and other batch processes in conjunction with Hyundai APIs.
  • Led a team of three artists and four engineers.
  • Real-time systems across data, game engine, and cloud.
  • Unity C# data ingestion, artist tools, and outputs to the AWS cloud, with Python data engineering and AWS cloud architecture underneath.
  • An AR Foundation car configurator, and an Unreal Engine car configurator with automotive shader exploration.
  • A C++ plugin for the creation of a lens distortion profile.

Impact & outcomes

  • 1,000,000+ configurations — a photoreal car configurator running in real time in a game engine.
  • Worked with Unity on their own automotive platform — which adopted more than a few of my ideas.

Stack Unity (C#) Unreal Engine AR Foundation Python data engineering AWS Cloud pixel streaming C++

  • Jan 2019 – Feb 2021.
  • A second, earlier engagement with the same company — Creative Technologist, Nov 2016 – Apr 2017 — sits on the Independent Era card as its own workstream.

Encore Hollywood

Senior Pipeline Developer — Episodic VFX

· Hollywood, CA

Episodic VFX at volume — pipeline scaling, compositing automation, and the first machine-learning prototypes in the shop.

  • Managed and scaled VFX pipelines supporting thousands of episodic shots across multiple concurrent productions.
  • Automated compositing workflows and render-farm optimization, improving throughput by 40%+.
  • Prototyped early ML integrations for asset tagging and pipeline intelligence.
  • A full TV episodic pipeline, from ingestion through delivery, running multiple shows per season.

Stack Compositing automation Render-farm optimization ML asset tagging

  • Jul 2017 – Aug 2018.
  • The profile puts a figure on the shot volume. The résumé says "thousands," and under D-017 rule 2a the résumé's wording is what ships — so no shot count appears on this card.

Independent Era

Technical Director — Contract Engineering

· Los Angeles, CA

Six years of contract technical direction — mocap and VFX pipeline, Nuke and Maya tool authoring, crowd work at volume, real-time prototyping, and the spatial-localization SDK that closes the period.

  • Tech Lead — Automotive Visualization (Digital Giant); Unity Developer — Spatial Localization SDK (Fantasmo); VFX / Real-Time TD (Timber, Royale, House of Moves, Illuminate Hollywood).
  • Developed a Unity C# SDK for visual-positioning localization at Fantasmo, letting a user localize and calibrate their physical position to within ten centimeters, alongside the research team in Munich.
  • Built a full motion-capture pipeline at House of Moves — ShotGrid and Tank API programming, MotionBuilder Python scripting, and a timelog manager in PyQt.
  • Handled a high volume of commercial crowd shots at A52 with Massive, using Nuke Python and standalone Python to manipulate massive data.

Impact & outcomes

  • Built AR on new localization technology at Fantasmo — the tech now known as VPS — years before Niantic shipped theirs.

HOUSE OF MOVES

Pipeline TD, full motion-capture pipeline. Jul – Dec 2012.

  • ShotGrid programming against the established API, and Tank API programming.
  • MotionBuilder Python scripting.
  • Timelog manager built with PyQt.

FREELANCE GENERALIST TD

Nuke, Maya, Houdini and Softimage tool authoring. Jan 2013 – Feb 2014.

  • Nuke (Python) — assembled custom V-Ray templates and gizmos, built a repair module for badly blended deinterlaced footage, and built show tools for environment, templates, gizmos and I/O.
  • Maya (Python, MEL) — implemented an Alembic toolset to speed up production, maintained V-Ray render-pass templates, and built a shader application for Alembic files out of Houdini using OpenMaya.
  • Houdini (Python) — built a show-specific FX environment, a particle debris OTL, and the I/O code to and from Maya.
  • Softimage (ICE) — lighting with Arnold.

ALMA MATER

Compositor and TD, stereoscopic feature. Nov 2013.

  • Nuke stereo compositing and ACES pipeline setup on The Lego Movie end credits.

A52

Massive TD, commercial crowd work. Aug – Sep 2015.

  • High volume of commercial crowd shots.
  • Nuke Python and standalone Python manipulation of massive data.

GENTLEMAN SCHOLAR

Compositor, commercial. Oct 2016.

  • Commercial compositing.

DIGITAL GIANT — CREATIVE TECHNOLOGIST

Rapid real-time prototyping. Nov 2016 – Apr 2017.

  • Rapid real-time prototyping.

FANTASMO

Unity developer, visual positioning system. Aug – Nov 2018.

  • Unity Developer — Spatial Localization SDK (Fantasmo).
  • A startup working on localization technology with a visual positioning system, where a user can localize and calibrate their physical position to within ten centimeters. Worked alongside the research team in Munich.
  • Rapid AR prototyping, and AR multiplayer for four players.
  • Developed the Unity C# SDK, which used a great deal of linear algebra.

Stack Nuke (Python) Maya (Python / MEL / OpenMaya) Houdini Softimage / ICE Arnold V-Ray Alembic Unity MotionBuilder ShotGrid / Tank API Massive RealFlow

  • Jul 2012 – Nov 2018, across seven engagements.
  • The résumé compresses these years into a single line; the profile is what supplies the individual shops, spans and titles below.

Timber

Nuke Compositor · TD

· Los Angeles, CA

A Nuke pipeline built from scratch and the compositing that ran on it — custom I/O nodes hooked into the studio's Python API, across commercial and music-video work.

  • VFX / Real-Time TD (Timber).
  • Nuke pipeline and compositing.
  • Custom Nuke setup with a variety of tools and workflows enabled — custom Read Image, Read Geo, Read Cam, Write Image and Write Geo, hooked into the pipeline Python API.
  • Compositing on the New Nintendo 2DS XL launch, and the fire and crowd shots on Kendrick Lamar's HUMBLE.

Impact & outcomes

  • Built the studio's Nuke pipeline from scratch — custom I/O hooked straight into the studio's Python API — and covered Flame Assist.

NUKE PIPELINE

Custom node set, hooked into the studio Python API.

  • Custom Read Image, Read Geo, Read Cam, Write Image and Write Geo nodes.
  • A custom Nuke setup with a variety of tools and workflows enabled, hooked into the pipeline Python API.

COMPOSITING

Commercial and music-video shots on the pipeline above.

  • Compositing on Introducing New Nintendo 2DS XL.
  • Fire shot and crowd scenes on Kendrick Lamar's HUMBLE.

Stack Nuke Python Custom Nuke gizmos Pipeline Python API Compositing

  • Nov 2015 – Apr 2017.
  • Freelance throughout two years.

Apple MARCOM

Technical Director — Contract

· Sunnyvale, CA

RealFlow water simulation for the Apple Watch keynote, and FX support for the staff around it — a summer inside Marketing Communications.

  • RealFlow liquid simulation, assisting staff with FX issues.

Stack RealFlow Liquid simulation FX support

  • Jul – Sep 2016. Three months, and the shortest era on the rail.

We Are Royale

Staff Technical Director · Creative Technologist

· Hollywood, CA

Two roles at one shop: Arnold lighting and pipeline integration first, then creative technology — AR and a live puppet system driven by real performers.

  • VFX / Real-Time TD (Royale).
  • Arnold lighting TD on Nike Evo, and Arnold XSI lighting with pipeline integration on Nike Tech Fleece.
  • Bathroom lighting and compositing on MTV Pulse, with Alembic integration across multiple 3D programs.
  • Bringing creative and development together across multiple mediums and platforms.
  • AR and puppet system in Unity 3D driving a character from two puppeteers and one motion capture actor.

Impact & outcomes

  • Built a live digital-twin puppet platform — a mocap actress plus two Henson puppeteers splitting the face — sent to Shanghai as the avatar of tennis star Li Na.
  • Unified the shop's pipeline across Arnold, XSI and Alembic.

STAFF TECHNICAL DIRECTOR

Lighting and pipeline. Feb 2014 – Mar 2015.

  • Arnold lighting TD on Nike Evo, and Arnold XSI lighting with pipeline integration on Nike Tech Fleece.
  • Bathroom lighting and compositing on MTV Pulse, with Alembic integration across multiple 3D programs.
  • Making things better and faster.

CREATIVE TECHNOLOGIST

AR and puppet systems in Unity. Mar – Aug 2015.

  • Bringing creative and development together across multiple mediums and platforms.
  • AR and puppet system in Unity 3D driving a character from two puppeteers and one motion capture actor — #AskLiNa.

Stack Arnold Softimage XSI Alembic Unity 3D AR Motion capture

  • Feb 2014 – Aug 2015.
  • The profile also nests a freelance generalist year (Jan 2013 – Feb 2014) under the same company; it sits on the Independent Era card, where its dates fall.

Sony Pictures Imageworks

Pipeline Developer

· Culver City, CA

Brought the studio's first Alembic geometry pipeline into a feature production, working with the format's original authors.

  • Integrated the studio's first Alembic geometry pipeline into the Hotel Transylvania production, working directly with Alembic's original authors.
  • Tool creation and support for the Layout department around the new Alembic technology.
  • Assisted animators with broken rigs and difficult technical issues, and supported the shop on Nuke 3D lighting questions.

Impact & outcomes

  • Worked on the studio's newly invented Alembic-based crowd technology, and drove Alembic into Layout for massive speed boosts.

Stack Alembic Geometry pipeline Maya (MEL / Python / PyQt) Nuke (Python) Katana (Python)

  • Nov 2011 – Jul 2012.
  • Front-end pipeline.

StereoD

Employee #5 — Stereo Conversion Pipeline

· Los Angeles, CA

Employee number five at a stereo-conversion startup, building the depth-conversion workflows the company then scaled on.

  • Pioneered the foundational stereo-from-depth conversion workflows alongside the scientist who invented the technique.
  • Joined as employee #5 as the company scaled from five people into one of the industry's leading stereo-conversion houses.
  • The pipeline went on to power conversions for major studio franchises.
  • As Stereoscopic Supervisor, responsible for designing a pipeline and tools for a stereoscopic production, along with bids, scheduling, and shot supervision.
  • Stereoscopic conversion on Avatar.
  • Multiple test demos — Iron Man, Transformers, The Rocky Horror Picture Show, and The Cabin in the Woods.

Impact & outcomes

  • Turned a just-invented depth-conversion technique into the production backbone that made a five-person startup worth buying — and Deluxe bought it.

Stack Stereo-from-depth conversion Production pipeline Shot supervision Bidding and scheduling

  • Acquired by Deluxe, 2011.
  • The profile records this engagement as Stereoscopic Supervisor, Aug 2009 – Apr 2010. The résumé's span and framing are what ship (D-017, rule 2a); the profile supplies the scope beneath it.

Early Post Era

Nuke Compositor · Lighting & Comp TD · Stereo Conversion

· Los Angeles, CA

The first three years — compositing and lighting on features and commercials, and the stereo-conversion work that led into StereoD.

  • Lighting and compositing technical director at Sony Pictures Imageworks on G-Force, responsible for the reconstruction of a second eye for use in a stereoscopic feature film.
  • Stereoscopic compositor at Prime Focus on the stereo conversion of Star Wars: Episode I.
  • Nuke pipeline TD at Illuminate Hollywood 3D — advancement and automation of Nuke tools, including a stereoscopic 2D-to-3D converter built for a two-reel feature film delivery.
  • Pipeline TD at Big Block, building studio infrastructure in Python, PyQt, Python OpenGL, MaxScript, .NET, Apache, CGI, JSON, XML and HTML, and helping the engineering department build a Deadline render farm expandable with rental nodes.

SWAY STUDIO

Nuke compositor. Oct 2008 – Jan 2009.

  • Nuke compositor at Sway Studio on a Quaker State commercial and BMW tests.

SONY PICTURES IMAGEWORKS

Lighting / compositing TD, G-Force. Mar – Jun 2009.

  • Responsible for the reconstruction of a second eye for use in a stereoscopic feature film.

SONY PICTURES ENTERTAINMENT

Nuke compositor / 3D generalist. Apr – Sep 2010.

  • Compositing and post-viz on Battle: Los Angeles.

COMEN VFX

Nuke compositor / 3D generalist. Sep – Nov 2010.

  • Full generalist duties from matchmoving to final output including stereoscopy, on The Fighter, Limitless, and Knights of Badassdom.

PRIME FOCUS

Stereoscopic compositor. Nov 2010 – Feb 2011.

  • Stereo conversion on Star Wars: Episode I.

ILLUMINATE HOLLYWOOD 3D

Nuke pipeline TD. Feb – Apr 2011.

  • Responsible for the advancement and automation of Nuke tools and maximizing efficiency in workflows and the pipeline.
  • Created a stereoscopic 2D-to-3D converter for use in a two-reel feature film show delivery.

INGENUITY ENGINE

Nuke compositor. Apr 2011.

  • Compositing commercial work.

BIG BLOCK

Pipeline TD and Nuke compositor. Jul – Nov 2011.

  • Built a flexible and expandable studio infrastructure using Python, PyQt, Python OpenGL, MaxScript, .NET, Apache, CGI, JSON, XML and HTML.
  • Helped the engineering department build and maintain a Deadline render farm, easily expandable with rental nodes.
  • Compositing for commercial, film and tests, while maintaining the facility pipeline infrastructure.

Stack Nuke Python PyQt Maya Stereo conversion Deadline render farm

  • Oct 2008 – Nov 2011, across eight engagements.
  • The second-eye reconstruction work on G-Force directly precedes the StereoD years.

Abstraction ladder

The same years, read as altitude rather than as employers.

  1. VFX GENERALIST
  2. ENGINEER
  3. TECHNICAL LEAD
  4. PRODUCT ARCHITECT
  5. SYSTEMS ARCHITECT
  6. COGNITIVE ARCHITECT

Technical expertise

Five domains, 58 capabilities, 79 cited cross-links. Every entry traces to a line of the résumé; the evidence for each one is the engagement it appears in above.

AI / MACHINE LEARNING

MODELS · AGENTS · SEMANTICS

  • MULTI-AGENT ORCHESTRATION core
    Runtimes that coordinate many models and tools toward one verified outcome. Custom frameworks, AutoGen
  • LLM INFERENCE OPTIMIZATION core
    Getting tokens out of a large model at a cost that closes. vLLM, INT4/AWQ, Speculative decoding, Routing gateways
  • MoE SERVING core
    Mixture-of-experts models at the trillion-parameter class, served in-house. vLLM, INT4/AWQ
  • GRAPH NEURAL NETWORKS exploring
    Learning over relationships rather than over rows. PyTorch Geometric, GAT
  • GRAPH-RAG & KNOWLEDGE GRAPHS active
    Retrieval that walks a graph instead of guessing from a flat index. Neo4j
  • VECTOR DATABASES core
    Approximate nearest neighbour at library scale. Qdrant, Pinecone, FAISS
  • OBJECT DETECTION & PERCEPTION active
    Finding things in frames, in real time, on a headset budget. YOLO
  • SYNTHETIC DATA & CORPUS ENGINEERING active
    Building the training set, then auditing it adversarially for leakage.
  • EVALUATION & ALIGNMENT core
    Deciding whether a model is good enough, with evidence rather than vibes.
  • LLM RELIABILITY ENGINEERING exploring
    Long-horizon execution that survives its own error rate.
  • CRISIS DETECTION core
    Multilingual detection with production canaries, failing closed.
  • JAILBREAK FIREWALL & MODERATION core
    The adversarial surface of a consumer product, treated as a system.

INFRASTRUCTURE & COMPUTE

GPU · CLOUD · DELIVERY

  • GPU SYSTEMS core
    Owned silicon — the reason inference economics are a design variable here. NVIDIA B200 Blackwell, RTX A6000, Multi-GPU clusters
  • TRITON KERNEL OPTIMIZATION exploring
    Where the throughput actually lives, once the easy wins are gone.
  • AWS SERVERLESS core
    Infrastructure as code, deployed as stacks, billed by the request. Lambda, CDK, DynamoDB, Cognito, CloudFront, WAF, ECS, S3, Bedrock
  • TERRAFORM active
    Declarative infrastructure outside the AWS-native path.
  • DOCKER & KUBERNETES active
    The unit of deployment for anything that is not a Lambda. Docker, Kubernetes
  • CI / CD core
    The gate that decides whether work is finished, running on a machine. GitHub Actions
  • REDIS STREAMS MESSAGING active
    How agents talk to each other without losing the message.
  • BACKEND & API DESIGN core
    Routes, contracts, and the boring parts that decide whether it scales.
  • INFERENCE UNIT ECONOMICS active
    Cost per token, measured on hardware you own, against the rented alternative.
  • AUTH & ENTITLEMENTS active
    Identity, session, and the server-side truth about what a user paid for. Cognito SRP, Google SSO (PKCE), Google Play Billing
  • RENDER-FARM ORCHESTRATION archived
    Scheduling thousands of frames across a finite pool of machines.
  • TELEMETRY PIPELINES active
    Getting field data back from deployments that are frequently offline.

GRAPHICS / VFX / REAL-TIME

RENDER · SIMULATION · COLOR

  • UNITY core
    The real-time engine most of the XR and visualization work is built in.
  • UNREAL ENGINE active
    The other real-time engine, when the render bar is the deciding factor.
  • XR / AR FOUNDATION core
    Tracking, anchoring, and the device abstraction underneath both.
  • SHADER DEVELOPMENT core
    Material and lighting authored at the instruction level. HLSL, Metal
  • PHYSICS & FLUID SIMULATION active
    Motion derived from forces rather than from keyframes.
  • COLOR SCIENCE core
    Perceptual colour as a measurable quantity — the bridge between film and ML. OKLAB, Film sensitometry, 3D LUTs
  • NUKE COMPOSITING active
    The node graph every VFX pipeline eventually has to speak to.
  • HOUDINI active
    Procedural FX, and the strongest argument for authoring geometry as a program.
  • MAYA / BLENDER PIPELINES active
    Asset authoring and the interchange between it and everything downstream.
  • PROCEDURAL GENERATION exploring
    Content described by rules, so it can be regenerated instead of stored.
  • REAL-TIME VISUALIZATION & DIGITAL TWINS active
    A live model of a physical system, rendered at interactive rates.
  • STEREO-FROM-DEPTH CONVERSION archived
    Turning a flat plate into two eyes, one depth decision at a time.
  • ALEMBIC GEOMETRY PIPELINE archived
    Baked geometry interchange — the format that made cross-vendor VFX possible.

VOICE & AUDIO

SYNTHESIS · PROSODY · PIPELINES

  • VOICE SYNTHESIS core
    Generating speech that a listener does not have to forgive.
  • TTS INFERENCE OPTIMIZATION core
    The latency budget between a keystroke and a spoken word. FP16, KV-cache
  • PROSODY OPTIMIZATION active
    Rhythm, stress, and intonation — where synthetic speech gives itself away.
  • NEURAL VOCODERS active
    The final stage that turns a spectrogram into a waveform.
  • iOS SPEECH FRAMEWORKS active
    On-device speech where the platform already ships the hard part.
  • AUDIO INTELLIGENCE PIPELINES active
    Treating audio as a searchable, classifiable signal rather than a file.
  • SPATIAL AUDIO active
    Sound positioned in a room the listener is standing in.
  • PRODUCTION VOICE SYSTEMS active
    Voice work as a delivered engagement, not a demo.

LANGUAGES & TOOLS

IMPLEMENTATION SURFACE

  • PYTHON core
    The ML and pipeline language; most of the serving and research code.
  • C# active
    Unity gameplay, tooling, and XR application code.
  • C++ active
    Engine-side and native plugin work where the frame budget is the constraint.
  • RUST exploring
    Systems work where a memory bug is not an acceptable failure mode.
  • DART core
    Client code, and the language this site's kernel is written in.
  • JAVASCRIPT / TYPESCRIPT core
    Web front ends and the serverless handlers behind them.
  • SWIFT active
    Native iOS plugins and platform framework integration.
  • SHELL / BASH active
    The glue that holds every build and deployment together.
  • FLUTTER core
    One client codebase across mobile and web — including this one.
  • PYTORCH core
    The research surface for everything trained rather than written. PyTorch Geometric
  • REACT active
    Conventional product front ends, when a canvas would be the wrong answer.
  • ASP.NET / VB.NET (LEGACY) archived
    Read-only fluency in the platform being migrated away from.

Research & development

Specifics available in conversation, under NDA where required. Figures on this panel are the résumé's measured figures; nothing here has been rounded up for the site.

SOVEREIGN INFERENCE BENCH

TRILLION-PARAMETER-CLASS MoE ON OWNED SILICON OPERATING

Trillion-parameter-class mixture-of-experts inference served from hardware in the building — quantized, benchmarked, and costed against the rented alternative.

SUSTAINED THROUGHPUT
1,100+ tok/s
PRIMARY GPU
NVIDIA B200 BLACKWELL ×4
AUX GPU
NVIDIA RTX A6000
QUANTIZATION
INT4 / AWQ
SERVING RUNTIME
vLLM
MODEL CLASS
TRILLION-PARAMETER MoE

Simulated profile. RELATIVE THROUGHPUT — ILLUSTRATIVE — the interactive version plots an illustrative curve here and labels it as such. No measured series exists for this panel, so none is shown.

MULTI-AGENT PRODUCTION SYSTEMS

ORCHESTRATION RUNTIME · ROUTING · VERIFICATION OPERATING

Agentic engineering infrastructure built and operated end to end, and used daily to ship real products rather than to demonstrate the idea.

RUNTIME
ORCHESTRATION + MODEL ROUTING
MESSAGING
COORDINATION BUS
VERIFICATION
EVIDENCE-ANCHORED AUDIT TRAILS
APPROVALS
HUMAN-IN-THE-LOOP GATES
DUTY CYCLE
USED DAILY IN PRODUCTION

LIVE CONSUMER AI PRODUCT

NAME WITHHELD — RÉSUMÉ DISCLOSURE LEVEL IN PRODUCTION

An LLM-powered companion app in production on AWS serverless, with a defense-in-depth safety stack that fails closed. The product is not named here; the résumé does not name it either.

CDK STACKS
17
PLATFORM
AWS SERVERLESS
SERVICES
COGNITO · DYNAMODB · CLOUDFRONT · WAF · SES
CLIENT
FLUTTER
AUTH
COGNITO SRP · GOOGLE SSO (PKCE)
BILLING
SERVER-AUTHORITATIVE ENTITLEMENTS
SAFETY
AGE GATE · CRISIS DETECTION · JAILBREAK FIREWALL

A.R.I.E.S.

AUTONOMOUS RESEARCH · SYNTHESIS · EXTRACTION INTERNAL R&D

A multi-agent autonomous research system that performs the literature review, so the operator argues with the synthesis instead of the corpus.

ARCHITECTURE
MULTI-AGENT AUTONOMOUS
FUNCTION
LITERATURE REVIEW
FUNCTION
SYNTHESIS
FUNCTION
KNOWLEDGE EXTRACTION

GNN CREATIVE-INTELLIGENCE RESEARCH

GRAPH LEARNING × PERCEPTUAL COLOR ACTIVE RESEARCH

Graph neural networks pointed at creative work — joint embedding spaces where perceptual colour is a first-class feature rather than metadata.

FRAMEWORK
PYTORCH GEOMETRIC
ARCHITECTURE
GAT
EMBEDDING
CLIP JOINT SPACE
COLOR MODEL
OKLAB · FILM SENSITOMETRY
CORPUS
ADVERSARIAL LEAKAGE AUDITED

Simulated profile. EMBEDDING SEPARATION — ILLUSTRATIVE — the interactive version plots an illustrative curve here and labels it as such. No measured series exists for this panel, so none is shown.

CLIENT DELIVERY — LEGACY MODERNIZATION

WEBFORMS → SERVERLESS, FIXED FEE, PHI IN SCOPE DELIVERED 2026

A twenty-year-old ASP.NET WebForms platform replaced with a serverless React SPA for a San Diego legal-services firm — delivered, paid in full, and trusted with a scoped autonomous-deploy authorization.

LAMBDAS
37
ROUTES
33
REPLACED
ASP.NET 3.5 WEBFORMS + SQL SERVER
DELIVERED AS
REACT / TYPESCRIPT SPA ON AWS
ENGAGEMENT
FIXED-FEE MILESTONE, PAID IN FULL
COMPLIANCE
HIPAA / PHI WORKSTREAM
AUTHORIZATION
SCOPED AUTONOMOUS DEPLOY

Education

Clients & brands

Infrastructure

Operates production ML infrastructure including 4× NVIDIA B200 Blackwell GPUs and RTX A6000 for large-model deployment, fine-tuning, quantized inference, and reliability research.

Contact

Direct mail reaches me fastest. Bring the problem rather than the pitch — inference economics, agent orchestration, or a pipeline that has outgrown its authors.

Availability. Configured, not live: open to staff and principal ML engineering roles, AI infrastructure leadership, and scoped Nimo Labs engagements.