AI / systems / research
she/her · they/them
Senior AI and systems engineer, founder, and independent researcher. Building AI systems since 2017, I work on agent infrastructure, developer tools, formal proofs, and scientific systems whose behavior can be inspected.
Flagship research · arXiv preprint
I authored a preprint giving a dimension-two counterexample to the separable Jacobian conjecture in characteristic two, with an explicit polynomial map, Jacobian determinant 1, separable degree 3, and a three-point fiber. I formalized the full result in Lean 4 / Mathlib with kernel-checked evidence.
Read arXiv:2608.02634 ↗I started taking computers apart and rebuilding them at seven, learning PHP and running a World of Warcraft phpBB forum at eleven, and using VMware at fourteen. I started deep learning work in 2017 with actor-critic agents in Atari games, separately studying Atari 2600 / 6502 assembly, then moved into GPT-era language models in 2019, including GPT-3 beta work and AI engineering at Knox Financial in 2020.
Architected the entire agentic system, ETL, and LangGraph workflow for Further’s AI homebuying product, working directly with the CTO on system design and implementation. Designed typed state and graph transitions, retrieval and tool orchestration, and integrated property/mortgage APIs with deterministic controls. At Knox Financial, I built housing search, distributed ingestion, PostgreSQL services, REST APIs, funnel instrumentation, and CI/CD controls.
I now build an agent-infrastructure startup and my own agent harness. My work spans multi-provider OpenClaw operations through Connie, Agent Pulse and SigilX, x402 payment admission, Axiom-RS, Ferrite’s Rust-first web framework, Ferrum’s JavaScript runtime, and supervision, observability, and memory tooling.
My working loop is simple: define the contract, build the system, inspect the evidence, test the failure paths, and describe the boundary honestly.
A focused set of public artifacts and clearly labeled independent prototypes across agents, runtimes, developer tools, and embodied research.
Building and testing a connectome-constrained adaptive controller: a 166,700-neuron MaleCNS v1.0 graph with 25.58 million signed sparse edges, DoOR-derived OCT/MCH inputs, dopamine-gated eligibility at 2,891 KC→MBON contacts, custom MLX/Metal CSR execution on Apple Silicon, frozen controls, telemetry, and a FlyGym/MuJoCo path toward closed-loop control.
Research implementation · adaptive systemsInteractive D–T tokamak research model with deterministic 0D plasma dynamics, explicit thermal and electric accounting, scenario controls, live diagnostics, JSON/CSV exports, and a Lean-checked ledger of exact accounting identities.
Research model · reviewable by designA Rust-first experiment toward a React-like UI runtime and Next.js-like application framework, with server rendering, file-system routing, client hydration, artifact builds, and TypeScript-facing APIs.
View repository ↗Local-first observability for Codex, OpenCode, and Claude Code sessions: event-driven activity state, provider adapters, hooks, redaction, metrics, and a browser dashboard.
View repository ↗Custom multi-provider agent operations through OpenClaw: persistent memory, scheduled workflows, scoped skills, queues, session isolation, liveness routing, payment admission, and proof-oriented services across OpenAI, Claude, MiniMax, Kimi, and Grok.
SigilX public skill ↗Cofounding a continuous, scalable distributed work system with Temporal-coordinated, PostgreSQL-authoritative runs, durable outbox start-or-join, fenced execution, cancellation sequencing, isolated Docker/QQ workers, broker-mediated model access, protected verification, persisted evidence, skills, plugins, and guardrails.
Independent systems · contract-firstBuilding Agent Cloud with my co-founder: a continuous, scalable distributed work system around useful-work state, Temporal orchestration, PostgreSQL authority, fenced execution, durable cancellation, audit and security controls, skills, plugins, guardrails, and evidence-led release decisions. The wider independent portfolio includes formal verification, Rust systems, and scientific computing research.
Architected the entire agentic system, ETL, and LangGraph workflow for Further’s AI homebuying product, working directly with the CTO on system design and implementation. Designed typed state and graph transitions, retrieval and tool orchestration, and integrated property/mortgage APIs with deterministic controls. Helped teammates adopt Cursor and contributed AI Genius UI improvements.
Built long-running elizaOS social-agent integrations across WebRTC and X/Twitter direct-message workflows, with persistent memory, action frameworks, and API-rate-limit handling.
Owned a CRM book of business in a technical-consultative B2B motion, covering discovery, solution design, customer engineering conversations, and closing across Verizon Business product surfaces.
Built GCP ingestion and automation feeding BigQuery and Vertex AI workflows; trained predictive models and monitored model and data quality for a live analytical product.
Built customer-facing housing search, distributed ingestion, PostgreSQL services, REST APIs, funnel instrumentation, and forecast visuals. Hardened CI/CD, reduced flaky-test noise, introduced deploy gates, reviewed backend code, and mentored junior engineers.
Scientific computing, formal methods, and security research connected by a practical interest in systems that accomplish clear goals.
Biophysical computation, connectome-constrained simulation, nuclear-fusion models, quantitative analysis, and machine-learning trading systems.
Research systems · Python / MLCode verification, exploit research, CTFs, prompt injection, and jailbreak evaluation, alongside the tools and review practices that make complex systems easier to understand.
Security research · evidence-ledBuilt typed local trading dashboards and agent-control surfaces around Hyperliquid and Polymarket data; trained Apple Silicon models through PyTorch MPS and native MLX paths. In TypeScript, I designed an observational e-graph market-consistency monitor with union-find classes, shared live-value subexpressions, residual scoring, and a real-time Next.js dashboard. A separate Rust/JavaScript context-pack prototype explores hash-chained evidence, deterministic typed-fact projection, dependency closure, direct-versus-optimized representation comparison, and fail-closed authority checks. Research records include a LightGBM 80-feature ladder over 435,279 out-of-sample rows and six purged walk-forward folds, plus first-passage Monte Carlo calibration over 796 OHLC windows.
Research tooling · MPS / MLX / e-graphsLet’s build something legible.
AI infrastructure, developer tools, formal verification, scientific computing, and technical product work.