Selected systems work

I work on architecture-heavy technical problems where failure has a concrete cost: latency, security, reliability, adversarial behavior, operational complexity, or a system that simply does not survive contact with its users.

My role is usually to define the system, build the critical path, and make the important constraints explicit enough that the architecture can be tested rather than merely discussed.

Signature case studies

Inference has to fit a hardware and latency budget

Real-Time Underwater Detection

Problem
Operational underwater detection that had to work outside lab conditions.
Constraints
Operational noise, simulation-to-real transfer, and sub-50 ms Jetson inference.
Architectural contribution
Led the hybrid DSP + ML architecture, simulation workflow, and edge deployment path.
Outcome
Sub-50 ms operational inference on embedded hardware
Operating envelope
Real-time underwater detection
latency
<50 ms end to end
hardware
Jetson, embedded
input
operational noise
training
simulation to real
tuned for
Pd against Pfa
  1. DSP features
  2. Deep CNN
  3. operator
binding constraint

Your AI has to run privately or offline

Secure Distributed ML Infrastructure

Problem
Enable distributed LLM work in disconnected, approval-heavy environments.
Constraints
Air-gapped operations, classified handling, and strict artifact approval.
Architectural contribution
Built the training, deployment, and operator workflow capability.
Outcome
Distributed fine-tuning operational across an air-gapped multi-GPU cluster
Read the case study
Operating envelope
Secure distributed ML
network
air-gapped
handling
classified, approval-gated
compute
Multi-GPU cluster
method
quantized adapters
span
data prep to operator
  1. data pipeline
  2. fine-tuning
  3. deployment
  4. operator
binding constraint

Participants have a reason to game the system

Adversarial Storage & Incentives

Problem
Make continued ciphertext storage provable rather than trusted.
Constraints
Adversarial participants, anti-precompute mechanics, and long-horizon incentives.
Architectural contribution
Architected challenge design, verification, reputation, and incentive alignment.
Outcome
Operated in production under continuous integrity proofs
Read the case study
Operating envelope
Adversarial storage
data
Ciphertext-only
peers
adversarial
challenges
seed-chained, unpredictable
commitments
Pedersen, per chunk
incentives
sliding window
  1. challenge
  2. proof
  3. verification
  4. incentive
binding constraint

Your agent workflow has to behave, across many customers

TurnkeyHQ

Problem
Turn an ambiguous vertical-AI thesis into an operable multi-tenant system.
Constraints
Tenant isolation, bounded model behavior, release evidence, and commercial controls.
Architectural contribution
Co-founded and carried the product through architecture, implementation, operations, and controls.
Outcome
14 services · 244 API routes · 75 agent skills · private beta
Read the case study
Operating envelope
TurnkeyHQ
tenancy
isolated per tenant
services
14
API routes
244
agent skills
75
stage
private beta
  1. channels
  2. deterministic plane
  3. bounded agent plane
binding constraint

Additional systems

RF Classification & Streaming ML

ML engineer, model development, production deployment · 2019

Prototyped an RF modulation classifier with +12pp accuracy improvement via attention-based spectral encoders, then deployed models and APIs into streaming production environments.

Structure Lab (GEX)

Architect, full-stack, quant engine design · 2026

An options payoff-engineering platform that uses mixed-integer linear programming to construct optimal multi-leg option structures from user-defined intent — replacing heuristic selection with constrained mathematical optimization.

Also: Agentic Data, SwarmOS, Capability Commons.

How I contribute

Architecture under ambiguity

Establish system boundaries, state, interfaces, constraints, failure modes, and the smallest architecture capable of proving the idea.

Critical-path engineering

Build or lead implementation of the components whose behavior actually determines whether the system works.

Technical strategy and de-risking

Translate architecture into evidence: benchmarks, threat models, observability, deployment boundaries, economics, rollback, and falsifiable decision criteria.

Best-fit problems

I am most useful when the architecture is not yet obvious, several technical or operational domains interact, incorrect assumptions become expensive later, or the system must survive real users, real infrastructure, or adversarial behavior.

If you already have a clean specification and simply need commodity implementation capacity, I may not be the highest-leverage fit.

Most of this work is bought as a bounded engagement: an architecture and risk review, a de-risking sprint, or ongoing fractional principal architecture.

See engagement types
Ask