ZebraFlux AI Character Engine
A dual-budget architecture for emotionally continuous virtual characters and companions.
Slow · explicit · reflective
Plans, explains and composes.
Fast · embodied · predictive
Anticipates, values and modulates.
Systems engineer · roboticist · AI researcher
Building at the seam between infrastructure, intelligence and physical machines — from production Linux to embodied AI.
CTO, ZebraFlux · 2026
Production Linux, Android platform work, robotics and machine intelligence. The common thread is systems that must operate under real constraints.
ZebraFlux—led by CEO Jonathan Nausley and CTO Maximus Sallam—is developing an AI character engine built around a fast, body-like evaluator and a slower reasoning system. The research target is continuity: characters whose behavior emerges from state, consequence and memory—not scripted dialogue trees.
A dual-budget architecture for emotionally continuous virtual characters and companions.
Slow · explicit · reflective
Plans, explains and composes.
Fast · embodied · predictive
Anticipates, values and modulates.
The fast posterior and slower anterior evaluate the same situation at different compute budgets. The working claim is that coupling—not two unrelated modules—is the source of unified behavior.
The posterior carries recurrent state and modulates the anterior through a learned opaque interface. The anterior experiences urgency and valence as bodily conditions rather than representing a separate “fast mind.”
Shuffled and deliberately deranged affect-vector controls test whether the language model actually uses the injected state. The frozen-model experiments failed that test, motivating joint from-scratch training.
The current gate is a compact joint run on Apple Silicon. Large rented-GPU training is deferred until the coupling produces measurable behavior at small scale.
The current research began as a practical engineering sequence: voice, embodiment, then a new theory of coupled cognition.
A Python conversational system built before the generative-AI wave: NLTK dialogue, speech recognition and Google WaveNet synthesis, with synchronized articulation.
A 130 cm robotic companion with TPE skin, a Raspberry Pi, two Arduino controllers and 3D-printed eye and mouth mechanisms—built for under $1,000.
A from-scratch dual-system research line testing whether self-awareness can emerge from coupling between shared evaluations at different compute budgets.
The work spans production infrastructure, Android platform engineering, machine-learning experiments and software that removes specific operational failure modes.
Built aosp_cf_arm64_only_phone-userdebug natively on an ARM64 Ubuntu host, then packaged working Cuttlefish device images and a glibc-native host package after resolving a musl-versus-glibc distribution-name collision.
Worked around the MPSGraph INT_MAX tensor limit by chunking the 50,257-token vocabulary projection instead of materializing the full batch-by-sequence logits tensor. A 51.4M-parameter benchmark reached 7,412 tokens per second while another MLX workload shared the GPU.
Progressed from a POMDP gridworld through a body-state environment and frozen-8B FiLM probes. A 120-record intervention battery found no causal response to numeric affect injection; failed v4/v5 gists redirected the program toward joint from-scratch training.
Building a provenance-preserving pipeline for acquisition, normalization, exact and MinHash deduplication, decontamination, quality filtering and tokenized mixture staging. The code slice uses Software Heritage identifiers so later licensing filters can be reconstructed.
Nine years at Nexcess/Liquid Web, advancing from Tier 1 through Tier 3 to Systems Administrator II across production hosting, incident response, ModSecurity and Python/Bash automation.
Operational discipline, hazardous-material response and systems thinking under consequence.
Field work across environmental consulting, renewable energy, construction and technical training.
Production infrastructure at scale, alongside enterprise information-security consulting through Phoenux.
Dorothy M1→M2→M3, from conversational software to embodied robotics to a research architecture for coupled cognition.
Leads the technical program behind the AI Character Engine alongside CEO Jonathan Nausley. Current work also includes service as a Wayne County Notary Public, driving with Uber and providing home care for Jonathan.
AAS in CIS—Cybersecurity in progress. Associate in General Studies, magna cum laude, completed in 2022. Dorothy M2 was developed as the final engineering project for ENGR-130.
The technical work sits inside a broader life shaped by military service, self-directed learning, caregiving, making and an insistence on understanding how things work.
Robotics, 3D printing and computing are parts of the same hands-on practice: study the system, build the mechanism, test it under constraints and revise what fails.
Outside that work, the recurring interests are space, gaming, shooting, hunting, DIY construction and reading. Learning is the reason for reading—not a separate pastime from the work.
I learned in 2022 that I have bipolar disorder. Earlier diagnoses included seasonal affective disorder, ADHD and emotional impairment. A bipolar diagnosis had been suggested before; I rejected it because I believed my behavior was reactionary and would stabilize if I were no longer antagonized. I was wrong.
I also believe I may be on the autism spectrum, though I have not explored that professionally. Devon Price’s Unmasking Autism opened my eyes to the possibility. That is a personal understanding, not a professional diagnosis.
Over the years I have developed strong skills for addressing and resolving mental-health-related issues quickly, without causing undue stress to others.
Service as a U.S. Army CBRN Defense Specialist was followed by environmental work, solar and construction, technical training, enterprise security consulting and nine years in production Linux infrastructure.
At Henry Ford College, I completed an Associate in General Studies magna cum laude in 2022 and continue toward an AAS in CIS—Cybersecurity. Dorothy M2 began there as the ENGR-130 final project. The current line extends that physical build into a computational theory of coupled cognition.
The original M1 voice system combines speech recognition, NLTK dialogue and Google WaveNet speech synthesis—the software layer later integrated with M2’s Raspberry Pi and Arduino-driven articulation. View repository
A controlled-network C2 lab with agent registration, hostname and tag identification, centralized task assignment, persistent IDs and automatic re-registration after server-state loss. View repository
A Magento 2 module that overrides Symfony Mailer and HttpClient behavior to stop forced ssl:// stream prefixes where STARTTLS or an upstream proxy owns TLS—without patching Composer-managed vendor files. View repository
A Puppeteer workflow for Project Honey Pot delisting on CentOS 7: discover the public IP, populate the request, pause for human CAPTCHA entry, retry bounded failures and optionally clean the working directory. View repository
A merged Android boot/recovery image toolkit restoring files removed from one upstream distribution. It ships unpack, repack and cleanup scripts plus Linux and macOS binaries for boot images, ramdisks, AVB and ChromeOS image handling. View repository
A LineageOS 21.0 device-support base for MT6735, MT6739, MT6755 and MT6763, organized around shared build configuration, device initialization, audio, sensors, thermal management and platform/vendor SELinux policy. View repository
Mobile appointments throughout Wayne County, with weekend availability. Meet at home, work, a hospital, care facility or another agreed public location.
Tell me where you are, how many signers are involved and how many notarizations the document requires. I will confirm the full fee before traveling.
Per act. Michigan law permits a notary to charge up to $10 for each notarial act.
Mobile rates are set in line with current Metro Detroit pricing. Your quote will separate the notarization fee from travel and any scheduling surcharge. See the Michigan Department of State fee guidance and a local market reference.
Acknowledgment, jurat, oath or other authorized act.
Travel from Dearborn Heights; final quote depends on distance and appointment requirements.
Saturday or Sunday, added to the agreed travel fee.
No charge for the first 15 minutes; any additional wait is agreed before it is incurred.
Confirming that a signer personally appeared, was identified and acknowledged signing a document voluntarily.
Administering an oath or affirmation for affidavits, verifications and other sworn documents signed in the notary’s presence.
Witnessing or attesting a signature when the document and requested notarial act allow it.
Certifying eligible copies when Michigan law permits; vital and public records must be obtained from the issuing agency.
Homes, offices, hospitals, care facilities and agreed public meeting places across Wayne County, subject to access and availability.
A notary verifies identity, willingness and the requested notarial act; a notary is not an attorney and cannot select the certificate, prepare documents or provide legal advice. Signers must personally appear with acceptable identification. Do not sign ahead of time when the requested act requires signing in the notary’s presence.
For mobile notary work, call directly or send a proposed date and time. Appointment requests are confirmed by phone or email.