← Dashboard

Senior AI Engineer · Anna, Texas

Kevin Savill

I build AI systems that connect to real workflows, from autonomous customer support to internal tools that take action.

Email · LinkedIn · GitHub

Showcase

Selected engineering work, from customer support and internal tools to connected AI agents.

Momentus · End-to-end system design and development

Autonomous support

Over 80% of support chats and tickets handled without human intervention.

Support chats and Zendesk tickets connect to the AI support system.
System overview. The outcome covers chat and tickets together.

I built a support automation system from the ground up, integrating Zendesk with a deployable customer chat widget. The work brought customer-facing chat and ticket automation into one support system.

What I built

  • End-to-end support automation integrated with Zendesk.
  • A deployable chat widget for customer support.
  • AI workflows that handle support conversations and tickets autonomously.

Measured outcome

More than 80% of support chats and tickets were handled autonomously, without human intervention, over the six months preceding September 28, 2026. The result combines support chat and tickets.

My contribution

I designed and built the system from the ground up, connecting the support platform, AI services, and customer-facing chat experience.

Zendesk · LLM orchestration · FastAPI · Chat integration · Systems integration

Independent project · Platform architecture and full-stack development

Koobin

AI agents, workflows, and tools under one permission model.

Koobin connects chat and workflows to governed agent tools and durable workers.
Architecture overview based on the Koobin source; not an application screenshot.

I built Koobin as a self-hosted platform for organizations to run AI agents against their business systems. It brings chat, visual workflows, retrieval libraries, connected tools, and usage accounting into one application.

Agents, workflows, and knowledge

  • Configure agents, model providers, tools, and structured outputs; review tool calls through approve, edit, or reject cards.
  • Build versioned visual workflows with branches, loops, parallel work, sub-workflows, agent calls, and human input.
  • Connect documentation libraries, remote MCP tools, and integrations including Slack, GitHub, Jira, Zendesk, and Rocketlane.

Access and visibility

  • Organization and department grants, OIDC single sign-on, MFA, encrypted credentials, and audit events govern access.
  • Track model and tool usage alongside agent reliability; durable tasks and schedules survive worker restarts.

My contribution

Platform architecture, full-stack development, workflow execution, tool integrations, and the controls that make agent activity reviewable.

AI agents · MCP · Django · React · PostgreSQL · Celery · SSO

Cytracom · Internal portal development and systems integration

Internal tools

A shared portal for support automation and team enablement.

Support teams use an internal portal connecting Zendesk automation, enablement tools, and operational APIs.
Capability overview based on the internal tools source; no internal screens or customer data are shown.

I built and expanded an internal tools portal that brought Zendesk support automation, communications administration, diagnostics, and everyday support utilities into one place. The goal was to help teams investigate issues, manage workflows, and complete repeatable tasks through usable tools.

Zendesk automation inside the portal

  • Ticket search and management, customer setup workflows, moderation, and spam management.
  • AI-assisted classification and response generation, with approval queues, escalation controls, and response history.
  • Agent availability, reassignment workflows, routing rules, scheduled alerts, and operational reporting.
  • Response evaluation, regression checks, documentation insights, and knowledge-review queues.

Tools for team enablement

  • Contact CSV creation, validation, conversion, and API utilities for desktop communications workflows.
  • Phone configuration tools for generating and modifying device settings, including multipurpose keys, line keys, and account registration.
  • Network monitoring views, email diagnostics, and phone-number lookups to support investigations.
  • Audio conversion, transcription, text-to-speech, and structured-data utilities for recurring support tasks.

One connected workspace

A React portal connects Python and Node.js services with Zendesk and communications APIs. Role-based access, shared navigation, audit views, and operational dashboards make the tools usable across support and enablement workflows.

My contribution

Portal and tool development, API integration, support automation, and translating hands-on support experience into workflows that other teams could use.

React · Python · Node.js · Zendesk · REST APIs · Internal portals · Team enablement

Projects

Coming soon · private source

Bit Universe

Build a town. Grow an empire across ten worlds.

Exa Metropolis at night in Bit Universe, showing the colony HUD, illuminated buildings, aircraft, and Build control.
Exa Metropolis on Earth, from the Bit Universe game website.

A 3D idle city builder with colony economies, building upgrades, vehicles, logistics, and a progression system spanning multiple worlds.

Begin with a small settlement on Earth and expand through 100 hand-made colony maps across ten worlds, ending on Triton. The game combines local city planning with an interplanetary economy: research upgrades, automate construction, and connect colonies through shared funds, ships, and trade.

Coming soon to iPhone and iPad, Android, PC via Steam, and VR. Release dates and supported VR headsets are still to be announced.

My role: Game design and development

  • Ten worlds with distinct systems: lunar solar power and batteries, Martian terraforming, floating Venus cities, finite asteroid reserves, and delayed income on Triton.
  • Governors, research, build queues, and automated upgrades develop a colony into a self-sustaining economy.
  • A shared fleet carries resources and construction supplies between worlds; trade routes connect exports with import demand.
  • Prestige Keys provide permanent income growth, with 41 achievements and upgradeable offline earnings of up to 48 hours.
  • Pure C# economy rules, evolving save migrations, and automated tests for economy and progression behavior.

Buildings that evolve

Building upgrades change the city itself, with scaffolding and new architectural stages. Players can lock a favorite style while continuing to progress. Vehicles, collectible coins, and illuminated windows bring the day/night cycle to life.

Five Moon habitat upgrade stages, from a small dome to a tall green-banded tower.
The Moon habitat progression, from dome to tower.

From one colony to a solar-system economy

Colony wallets grow into world funds and an Interworld Treasury. Planet-specific resources, fleet fuel, market events, and cross-world deliveries connect individual cities into a larger simulation.

Bit Universe system map showing Earth, the Moon, Mars, Venus, the asteroid belt, and outer-planet destinations through Triton.
The in-game system map and the Far Out achievement.

Unity · C# · Simulation · NUnit

Visit Bit Universe · Explore the worlds · Player guide

In development · private source

Non-Euclidean Neon Arena

Dodgeball, Laser Tag, and Tracer in impossible nightclubs.

Non-Euclidean Neon Arena development build with neon courts and robot players behind the main menu.
Actual development build, version 0.27.5.

A robot arcade-sports game where portals, warped scale, and gravity-changing spaces turn the arena into part of the strategy. Play Dodgeball, Laser Tag, or Tracer with bots, friends, or local split-screen players.

Built in Unity and C#, the project brings together impossible-space rendering, host-authoritative multiplayer, portal-aware opponents, a robot editor, and a continuous original soundtrack.

My role: Game design and development

  • Rendered folds connect spaces that cannot physically fit together: larger interiors, stretched rooms, mirrored routes, and inverted gravity.
  • Three sports share courts, local views, bots, and multiplayer rules; projectiles and beams interact with the arena folds.
  • Bots navigate portals, remember pickups, predict shots, and time defensive actions.
  • Career statistics, a trick book, and earned robot cosmetics reward play without changing competitive abilities.
  • Desktop packaging for Windows, macOS, and Linux; shared keyboard, controller, and touch input paths.

The arena is the mechanic

A shortcut can change scale, reverse an angle, or lead into a room larger than its exterior. The challenge is making these transformations agree across rendering, movement, projectiles, bot navigation, and networked play.

A doorway into the larger neon interior of the Police Box court.
Development capture of the Police Box court; the interior is larger than the exterior.

Unity · C# · URP · Game AI · Multiplayer

In development

Ram Ramp

Keep the ram aboard. Brake, launch, and chain the flight.

A toy ram flies toward a golden ring above a farm course with distance and speed indicators.
Development capture of Ram Ramp in the portrait phone layout.

A physics-driven arcade game about a toy ram riding in a pickup. Balance the throttle to keep it aboard, brake at the right moment, then steer through farmyard bounce chains.

Designed around short mobile sessions, with a portrait interface, guided first-flight lessons, responsive desktop controls, and an offline progression system.

My role: Game design and development

  • Truck acceleration, cargo friction, braking, and ram flight combine into a skill-based launch rather than a single power meter.
  • Twenty-two campaign stages span farm, quarry, and snow courses, with trampolines, updrafts, ice lanes, and earned flight abilities.
  • Long Haul chains seeded course sections; personal-best ghosts, Stunt Cards, missions, and trophies add replay goals.
  • Original procedural toy models and synthesized reactive audio, with cosmetic customization and atomic local saves.
  • Windows and Android build pipelines; physical-phone validation and storefront release work remain in progress.

Unity · C# · Physics · Mobile · Procedural art

In development

Scriptlands

Survive, build a factory, and program every machine.

Scriptlands machine editor running a pull, grind, and push script for a Macerator.
Development capture of the in-game machine script editor.

A third-person voxel survival and automation game where programming is part of the technology progression. Mine, craft, and explore, then connect machines, pipes, power, and signals with scripts written in a Python subset.

The playable slice includes procedural terrain, crafting, farming, night enemies, scripted machines, a quest book, and direct-IP co-op. Early syntax tiers are implemented; the broader language and technology progression remain in development.

My role: Game design and development

  • A custom interpreter runs player scripts inside a bounded operation budget, separate from the Unity presentation.
  • An in-game editor supports running, stepping, line tracing, and readable errors for factory scripts.
  • Pure C# world and factory simulation covers item pipes, generators, electric machines, batteries, and signal networks.
  • Server-authoritative co-op uses a custom TCP protocol, shared world rules, and server-checked movement.
  • Seeded biomes, caves, ore, world saves, and adaptive synthesized music connect exploration with factory building.

A world built around automation

The factory is part of a survival world. Players start with tools and hand crafting, repair a broken machine near spawn, and unlock the syntax that makes more complex production possible.

Procedural voxel forest terrain in Scriptlands.
Development terrain capture from the Scriptlands source checkout.

Unity · C# · Custom interpreter · Voxel worlds · Co-op

In development

Elsewake

Rebuild a homestead where other histories break through.

Elsewake loop: explore the lakes country, recover supplies, rebuild a homestead, and enter the Wake.
Gameplay systems overview drawn from project documentation; not a gameplay screenshot.

An open-world survival game set in a lakes region where fragments of other histories appear at night. Scavenge by day, build a home, and enter the Wake quietly: noise can tear a fragment apart and release echo infected.

The project combines exploration and a persistent homestead with interconnected survival, vehicles, settlement progression, and a streamed world. The authored expedition path reaches the Ridgeway and greenhouse beats; co-op remains experimental.

My role: Game design and development

  • Nightly Wake incursions connect stealth, sound, rare supplies, and recordings from alternate histories.
  • A homestead grows through recovered components for water, radio, medical support, power, and vehicles.
  • Interlocking systems include weapon condition, ammunition, food spoilage, radiation protection, and vehicle fuel and components.
  • Streamed locations, settlements, contracts, and configurable survival rules support different ways to explore.
  • Unity/C# development with persistence, world streaming, and experimental host-authoritative co-op.

Unity · C# · World streaming · Survival systems · Persistence

In development · self-hosted

Voice Identity Studio

Turn recordings into reviewed, reusable speaker identities.

Voice Identity Studio workflow: ingest media, identify speakers, review and deliver.
V4 workflow overview; schematic rather than an application screenshot.

A self-hosted audio studio for vocal isolation, speaker diarization, transcription, and matching speakers across recordings, with a browser-based segment editor and per-speaker exports.

The V4 application extends the earlier processing pipeline into a complete review workflow: resumable media ingestion, identity matching, human review, voice profiles, and voice-replaced renders. Local GPU engines are the default; external providers are optional.

My role: Full-stack development and audio pipeline engineering

  • Match speakers across files with confidence scores, review queues, project or workspace scope, and reversible identity merges.
  • Review transcripts and speaker segments in a timeline editor, audition original and separated stems, and export speaker audio.
  • Resumable uploads support browser-side audio extraction or retained video proxies.
  • Consent-gated voice profiles and voice replacement retain provenance; native RVC supports training and conversion.
  • Face tracking in retained video can suggest links to voice identities; ambiguous matches remain for review.
  • FastAPI and React coordinate PostgreSQL-backed state and background GPU jobs, with authenticated API and MCP access.

Python · FastAPI · React · TypeScript · PostgreSQL · GPU processing

In development · self-hosted

Koobin

AI agents, workflows, and tools under one permission model.

Koobin connects chat and workflows to governed agent tools and durable workers.
Architecture overview based on the Koobin source; not an application screenshot.

A self-hosted platform for organizations to run AI agents against their business systems. Koobin combines streaming chat, visual workflows, retrieval libraries, realtime voice, and usage accounting with explicit access grants and tool approvals.

The architecture separates the web application from durable background work, so long-running agent turns, media processing, and scheduled tasks have their own worker queues.

My role: Platform architecture and full-stack development

  • Configure agents, model providers, tools, and structured outputs; review tool calls through approve, edit, or reject cards.
  • Build versioned visual workflows with branches, loops, parallel work, sub-workflows, agent calls, and human input.
  • Connect documentation libraries, remote MCP tools, and integrations including Slack, GitHub, Jira, Zendesk, and Rocketlane.
  • Organization and department grants, OIDC single sign-on, MFA, encrypted credentials, and audit events govern access.
  • Track model and tool usage alongside agent reliability; durable tasks and schedules survive worker restarts.

Django · React · Celery · PostgreSQL · Redis · MCP

Live · free to use

Universal Flow Puzzle Solver

From a puzzle screenshot to a solution, entirely on your device.

Before-and-after view of a 35-region irregular puzzle with cutouts. Four colored paths connect matching endpoints and cover every region.
Actual imported puzzle geometry and independently verified solution from the full Python solver. Irregular-region solving is not yet available in the live browser edition.

A free web app that turns a screenshot of a Flow / Numberlink-style puzzle into a solved board. Choose, drop, or paste an image, check the detected colored dots, correct any mistakes, and solve directly in the browser.

Image processing and solving run on the visitor's device. Puzzle screenshots are not uploaded to a processing server. The app is hosted on Cloudflare Workers Static Assets using free static hosting, with no backend compute service required.

My role: Full-stack development and solver engineering

  • Screenshot detection finds regular board outlines and color pairs, with a visual editor for correcting dots and adjusting board corners.
  • Z3 compiled to WebAssembly solves path constraints locally, removes disconnected cycles, and independently checks the resulting solution.
  • Dedicated Web Workers keep image processing and search off the interface thread, with progress details and a Stop control.
  • An IndexedDB library saves puzzles in the browser; .flow import and export let users move their puzzles between devices.
  • The live edition supports square-grid boards up to 20 × 20, including rectangular layouts and manually entered holes. Bridges, walls, and warps are not supported yet.

Solving beyond regular grids

The full solver models a puzzle as connected regions rather than assuming a rectangular grid. This example uses 35 irregular cells with cutouts: all four pairs are connected, every region is covered, and the resulting paths pass an independent validity check.

  • Example: IMG_3246 from the source puzzle corpus, solved with the Python Z3 engine.
  • The hosted device-only app currently supports square-grid boards, including rectangular layouts and manually entered holes.

Built for static hosting

The browser downloads a compressed solver runtime, verifies its size and integrity, and initializes it locally. React and TypeScript provide the screenshot-to-solution workflow, while Cloudflare serves the app and WebAssembly assets. No account is needed to solve a puzzle.

The live browser experience

The public site keeps screenshot processing, board review, and square-grid solving on the visitor's device.

Flow Puzzle Solver displaying a solved five-by-five demo board, an on-device processing notice, and an independently checked solution.
The live site solving its built-in demo in the browser, captured September 30, 2026.

React · TypeScript · Z3 · WebAssembly · Web Workers · IndexedDB · Cloudflare

Try the puzzle solver · Usage guide

Public source · experimental

Pong reinforcement learning

Training and evaluating an agent through play.

Reinforcement learning loop connects observations, agent actions, environment, and rewards.
Training-loop overview; no benchmark score is implied.

A reinforcement-learning project for Atari Pong, with a DQN training pipeline and tools for evaluation and interactive play.

My role: Applied machine learning

  • Replay-buffer and target-network integration in the training loop.
  • Configurable exploration, checkpointing, and evaluation.
  • Training metrics and reproducible experiment configuration.

PyTorch · DQN · Gymnasium · TensorBoard

View training source

Public source · experimental

Digit recognition

Draw a digit and explore a convolutional network.

Digit recognition: drawing input, convolutional features, classifier, digit prediction.
Model and interface overview; schematic, not a live prediction.

A handwritten-digit recognition project with a convolutional neural network, training tools, and both web and desktop drawing interfaces.

My role: Applied machine learning and interface development

  • Convolution, batch normalization, pooling, and dropout in the model.
  • Training augmentation, learning-rate scheduling, and early stopping.
  • A web drawing canvas with predictions and convolution-layer activation previews.

PyTorch · CNN · Flask · Tkinter

View model and app

Personal website

kevinsavill.com

A portfolio you navigate like a 2008 console dashboard.

Kevin Savill's console-inspired dashboard with the Showcase tile selected.
The portfolio dashboard, captured in a desktop browser.

A personal site inspired by the Xbox 360 New Xbox Experience dashboard, built with plain HTML, CSS, and JavaScript.

My role: Design and development

  • One navigation model for keyboard, mouse, touch, and controller inputs.
  • Shareable detail panels and browser-back navigation.
  • Synthesized interface sounds and an animated avatar.
  • A static reading view and downloadable résumé alongside the dashboard.

HTML · CSS 3D transforms · JavaScript · Web Audio · Gamepad API

Open reading view

UT Dallas CS capstone

Virtual chatbot

Computer Science capstone at UT Dallas.

A team capstone project completed for my B.S. in Computer Science at the University of Texas at Dallas. We combined a React and Next.js interface, a Rasa conversational engine, and a Flask backend to connect conversations with application data.

My role: Capstone team member

  • REST integrations with databases and external APIs.
  • Context-aware conversational responses.
  • In-house two-factor authentication.

React · Next.js · Rasa · Flask · REST APIs

Personal project

Plants vs. Zombies automation

Image recognition turns game screens into automated actions.

A screen automation bot for Plants vs. Zombies minigames and the Zen Garden, using OpenCV image matching and PyAutoGUI controls.

My role: Automation and interface development

  • DearPyGui controls for selecting a mode and run duration.
  • Slot-machine minigame automation for collecting sun and placing plants.
  • Zen Garden routines for plant care, resource purchases, and snail feeding.

Python · OpenCV · PyAutoGUI · DearPyGui

Experience

Ntracts

100% remote

Developing AI capabilities for internal operations and product experiences.

  • Building internal AI tools and automation to support teams and business workflows.
  • Contributing to product capabilities through AI engineering and systems integration.

Senior AI Engineer

Jun 2026 – Present

Advance Velocity

100% remote

AI consulting focused on support automation, connected systems, and customer-facing AI experiences.

  • Built an end-to-end Zendesk support automation system and deployable chat widget for Momentus.
  • Over 80% of support chats and tickets handled autonomously without human intervention, across the six months preceding September 28, 2026.

Senior AI Consultant

Jun 2026 – Present

AI Consultant

Mar 2025 – Jun 2026

Cytracom

Allen & McKinney, Texas

Business communications provider. Five roles in five and a half years, from Tier 1 support to solutions engineering.

  • Rebuilt internal portals with integrated public and internal APIs, substantially improving organizational efficiency
  • Built a shared internal portal for Zendesk automation, diagnostics, and team enablement.
  • Developed ticket moderation, routing, reporting, and communications administration tools.
  • Applied agentic workflows to code-development automation.
  • Built a GUI for parsing and generating Grandstream, Yealink, SNOM, and Polycom phone configurations.

Solution Engineer

Jan 2025 – Jun 2026 · McKinney, Texas · On-site

Automation Engineer

Apr 2024 – Mar 2025 · Allen, Texas · Hybrid

  • Built and maintained AI-automated solutions for support ticketing
  • Developed and maintained internal support tools
  • Led resolution of bug-related and engineering issues
  • Managed technical programs and monitored ICU boards

Senior Technical Support Engineer

Apr 2022 – Apr 2024 · Allen, Texas

  • Resolved complex bug-related and engineering issues for the Application Support team
  • Maintained availability and reliability of published internal support tools
  • Configured call routing, call-quality monitoring, and BLF features, including firmware upgrades
  • Supported desktop, mobile, and integration platforms
  • Ran at least one feature or bug demo per week to train the team

Technical Support Tier 2

Oct 2021 – Apr 2022

Technical Support Tier 1

Dec 2020 – Oct 2021 · Allen, Texas

Skills

Languages

Python · TypeScript · JavaScript · C# · Java · SQL · HTML & CSS

AI & Automation

LLM orchestration · AI agents · MCP · Retrieval-augmented generation (RAG) · Hybrid semantic search · Visual workflow engines · Human approval workflows · Agent evaluation · Usage metering

Machine Learning & Vision

PyTorch · scikit-learn · OpenCV · MediaPipe · Object tracking · Head-pose estimation · Reinforcement learning · Gymnasium · Stable-Baselines3 · MuJoCo robotics simulation

Audio & Media

Speaker diarization · Speech transcription · Speaker embeddings · Vocal separation · RVC voice conversion · GPU processing pipelines · FFmpeg · Media provenance

Web & APIs

React · Next.js · FastAPI · Django REST Framework · Flask · Node.js / Express · REST APIs · WebSockets · WebRTC integration · Web Workers · WebAssembly integration

Data & Background Jobs

PostgreSQL · pgvector · Redis · Celery · SQLAlchemy · Database migrations · S3-compatible storage · IndexedDB

Games & Solvers

Unity / URP · UI Toolkit · Procedural terrain and meshes · Physics gameplay · Game AI · Authoritative multiplayer · Custom scripting interpreters · Economy simulation · Save migrations · Procedural audio · Z3 constraint solving · Graph modeling

Security & Access

OIDC / SSO integration · Multi-factor authentication · Multi-tenant permissions · Credential encryption · Audit logging

Delivery & Testing

Docker / Compose · Linux · Cloudflare static hosting · Git · pytest · Playwright · Vitest · NUnit / Unity Test Framework · Deterministic replay and validation

Systems & Tools

Systems engineering · Requirements engineering · Systems integration · Zendesk API · Microsoft Azure · .NET / WPF · USB HID integration · VoIP configuration tooling

Education

M.S. Systems Engineering and Management

The University of Texas at Dallas
Jan 2025 – May 2028 (expected)

B.S. Computer Science

The University of Texas at Dallas
Aug 2022 – May 2024

Associate’s degree, General Studies

Collin College
Aug 2020 – Dec 2021

Credentials earned