Skills & Capabilities
45 skills across 8 categories — from secure networking and embedded systems to production machine learning and agent infrastructure.
AI & Agent Infrastructure
Autonomous systems on hosted and self-hosted models
Claude Agent SDK
expertProduction agent pipelines with tool use, sub-agents, and sandboxing.
LangGraph
expertStateGraph orchestration, human-in-the-loop interrupts, checkpoint-resume after a restart.
Anthropic API
advancedDirect API integration, prompt engineering, tool use, batching.
Claude Code
expertDaily-driver agentic coding; custom skills, hooks, MCP servers.
MCP Clients & Servers
expertModel Context Protocol integrations — building servers, wiring clients, designing tool surfaces.
OpenAI API
advancedChat Completions, function calling, structured outputs, embeddings.
OpenClaw
advanced6 deployments across self-hosted infra; custom skills, hooks, and tooling; upstream PR contributor.
Self-Hosted LLM Serving
advancedOpen-weight vision models on owned GPUs behind a load balancer (vLLM), so data never leaves the building.
Structured Output
advancedSchema-bound, constrained replies that turn a large model into a fast classifier you can check.
Multimodal Document AI
advancedEmail, attachments and scans read by a vision model, with OCR and PII redaction before anything is used for training.
Applied ML & Evaluation
Pricing models and LLM pipelines, measured before they ship
Quantile Regression (LightGBM)
advancedPrice models that return a range instead of a point: low, typical and high, one model per segment.
Backtesting
advancedTime-split rolling backtests with monthly refits and paired bootstrap intervals against the incumbent benchmark.
Conformal Calibration
advancedRanges widened until the stated share of real outcomes falls inside them.
Win-Probability Modelling
advancedLogistic regression with partial pooling per customer, compared against gradient boosting.
LLM Evaluation Harnesses
advancedStratified holdouts, per-category precision and recall, calibration error, and checks against what the business actually did.
Shadow Deployment
advancedNew models score every request in shadow and compete per segment before they are allowed to change a price.
Backend & Systems
Production APIs, persistence, orchestration
Java + Spring Boot
expertProduction orchestration engines, Hibernate ORM, REST APIs.
Python + FastAPI
advancedAsync-first APIs, dependency injection, OpenAPI docs.
PostgreSQL + Hibernate
advancedSchema design, migrations and query tuning on databases with millions of rows.
Docker
expertMulti-stage builds, Compose, and Docker-in-Docker sandboxes with dropped capabilities and network isolation.
Go
advancedCommand-line and deploy tooling.
Architecture Guard Tests
advancedModule boundaries, permissions and migration safety enforced by tests that fail the build.
Zero-Downtime Migrations
advancedExpand-and-contract schema changes with Flyway; destructive changes are rejected automatically.
Testcontainers
advancedIntegration tests against a real PostgreSQL, cloned per test from a template.
Platform & Reliability
Deploys, observability and access control
Blue/Green & Canary Deploys
advancedMerge-to-deploy with a boot gate, smoke test, canary window and automatic rollback.
Observability
advancedPrometheus, Grafana, Alertmanager and Loki; unit-tested alert rules and GPU monitoring.
Identity & Access
advancedOIDC single sign-on, scoped API keys, signed service-to-service calls, audit logs.
Credential Isolation for Agents
expertProxies that hold the tokens, so a sandboxed agent can push code without ever seeing a secret.
Infrastructure as Code
intermediateOpenTofu and cloud-init provisioning, image scanning, Cloudflare tunnels and DNS.
Networking & Secure Systems
Production VPN, network stack integration, cryptography
VPN Architecture
expertMulti-path routing, global relay architectures, high-assurance deployments.
Network Stack Integration
expertWintun (Windows TUN), Network Extension (Apple), VpnService (Android).
Cryptography Fundamentals
advancedApplied crypto for secure routing; collaboration with dedicated cryptography teams.
TUN/Driver Development
expertNative C drivers for layer-3 TUN, kernel-adjacent code paths.
Cross-Platform Engineering
Native iOS, macOS, Android, Windows, Linux
iOS / Swift / Network Extension
advancedNative VPN clients, Apple permission handling, system integration.
Android / Kotlin / JNI
advancedVpnService implementations bridging C networking via JNI.
Windows / Native C
expertVisual Studio C/C++, Wintun layer-3 TUN, GUI applications.
Linux Systems
expertProduction deployment and troubleshooting in restricted-access sites worldwide.
Frontend & Visualization
Modern React stacks and creative web experiences
React / Next.js
advancedApp Router, Server Components, streaming SSR, modern patterns.
Three.js / R3F
intermediateParticle fields, glow orbs, scene composition with React Three Fiber.
Framer Motion
intermediateScroll-triggered animations, stagger sequences, gesture handling.
TypeScript
advancedStrict mode, generic types, discriminated unions, type-driven design.
Embedded & Hardware
Firmware, FPGA, board bring-up, BSP customization
Embedded C
expertProduction firmware on commercial industrial control systems.
Buildroot
advancedCustom Linux BSPs for airborne radio platforms.
Xilinx FPGA (Vivado / XSCT)
intermediateFPGA integration on dev boards and target hardware.
JTAG Bring-Up
advancedBoard bring-up on factory-assembled units; hardware validation.
How I work
Six habits behind the skills above. Each one is visible in the open-source embry0 repository or in client work.
Prove a check can fail
Before I rely on a test, an alert or a gate, I make it fail on purpose. A check that has never been red proves nothing.
Write the finish line first
A job starts with a short list of checkable conditions. The agent builds to that list, and review and QA judge against the same one.
Shadow before live
A new model scores real requests beside the old one, and takes over a segment only after it has won there.
Record the decision
Every non-obvious choice gets a short record: what was chosen, what was rejected and why. embry0 has thirty of them.
Say what does not hold
Security notes list the guarantees and, beside them, the cases where they do not apply.
Verify like an operator
A change is done when I have driven it through the real interface and confirmed which revision is being served.