AI and agent-systems engineer with a security focus and a pure-mathematics foundation. I build and ship LLM-based agentic systems end to end.
I design multi-agent LLM systems, the data and security layers underneath them, and the infrastructure that deploys them, including to air-gapped environments. I treat security as a design property rather than a bolt-on, and I keep every claim matched to what the code does: the repositories below run, are tested, and report real numbers from actually running.
Four interoperating projects that simulate a satellite constellation's ground segment: the data plane, the brains, the shield, and the deploy system. One coherent system. liftoff deploys the stack that groundstation operates over constellation's telemetry, with aegis authenticating the command path. Each project also stands alone and runs on its own.
| Project | What it is | Real headline number |
|---|---|---|
| constellation | Telemetry data plane: streaming ingestion (Redpanda → TimescaleDB), unsupervised anomaly detection, and a React/TypeScript ops console | Anomaly detection F1 0.720 (recall 0.808) on held-out satellites, over 180,000 streamed telemetry records |
| groundstation | Agentic mission-ops copilot: a LangGraph multi-agent graph operating the system through MCP tools and RAG-cited playbooks, with tiered model routing and a human-in-the-loop approval gate | Trained anomaly-type classifier 96.7% accuracy, 0.931 macro-F1 (held-out satellites) |
| aegis | Security layer: Ed25519 signed-command authentication (replay protection, key rotation, verifying chain), mutual TLS, and an adversarial harness with a measured defense stack | 19-case corpus, 19/19 succeed undefended and 0/19 defended. Eight cases were written by reading each guard and going around it: seven worked, dropping the block rate to 63% before the gaps were closed |
| liftoff | Deploy and build-reliability layer (Go): one declarative suite spec rendered for cloud, on-prem, and air-gapped bare-metal, with SLO-driven rollback | All three targets pass terraform, helm/kubeconform, and shellcheck, with byte-for-byte deterministic output |
- athena-ai: a deployed real-vs-AI image classifier (EfficientNet-B0) served behind an API with a live demo: 0.996 accuracy and 0.9999 ROC-AUC on held-out data, a verdict in roughly 150 ms, and an honest model card.
- slew: a 3-DOF spacecraft attitude-control simulator in C++20: quaternion-feedback control over real rigid-body dynamics (Euler's equations) with a saturating actuator. A default 30-degree slew settles to 0.0008-degree final error at a 1 kHz control rate, with a bit-identical trajectory every run. An optional four-wheel reaction-wheel model exchanges momentum instead of inventing torque, conserving inertial angular momentum to 8.4e-6 of the momentum exchanged.
- Languages: Python, Go, C++, TypeScript/JavaScript, SQL
- AI / agents: LangGraph multi-agent orchestration, MCP (servers and clients), RAG over a vector DB (pgvector), tiered Anthropic Claude model routing, PyTorch model training and evaluation
- Data: Kafka / Redpanda, stream processing, TimescaleDB and time-series, anomaly detection, pandas / NumPy / scikit-learn
- Security: prompt-injection / jailbreak / exfiltration red-teaming, Ed25519 signed command authentication, mutual TLS, replay protection, threat modeling
- Infrastructure: Terraform (EKS + GPU node groups), Helm, Kubernetes CRDs and controller-runtime operators, Docker, GitHub Actions, air-gapped and GPU deployment