From native apps and encrypted APIs to local-model routing and tool-using agents, I engineer the whole authority chain: model, tools, data, policy, observability, and recovery.
Hermes, my independent agent lab, runs through one loop: an orchestrating agent with explicitly scoped tools, durable plain-text memory it reads before working and writes back to after, and a local model for small, verifiable tasks. It plans, builds, checks, remembers, and recovers without receiving ambient authority.
The orchestrator. Plans work, calls scoped tools, and writes and reviews code under my supervision.
Typed tool boundaries between the model and the code, data, and services it may use.
Durable memory. Plain Markdown, versioned, human-reviewed before anything becomes canon.
On-device inference for quick work whose answer can be verified at a glance.
My independent portfolio spans React and TypeScript web applications, Swift native clients, Rust desktop software, edge workers, PostgreSQL services, authentication, encryption, deployment, and CI. The model is one subsystem. The engineering around it determines whether the product deserves trust.
At app.wadevo.com, I built a full-stack personal-finance product with cash-flow forecasting, bill and credit tracking, and natural-language financial search. AI sits beside a deterministic attention engine that keeps urgent calculations fresh without waiting on a model.
The surrounding system is the achievement: invite-only authentication, tenant-scoped data, authenticated encryption for sensitive fields, CSRF and rate limiting, moderation, prompt-injection defenses, and tamper-evident audit logs.
These are the repeatable behaviors behind my agent work. They are designed to travel without carrying anyone else's data, infrastructure details, or institutional context with them.
/orientReads the smallest relevant context set and separates verified state from assumptions before planning./reviewClassifies scope and risk before implementation, then returns a clear proceed, caution, or stop verdict./routeChooses local or frontier inference according to sensitivity, context needs, and verification cost./researchBuilds an answer from authoritative sources with visible provenance and bounded claims./verifyTests the evidence behind a claim before allowing it to become status, memory, or authority./handoffWrites decisions, deliverables, open risks, and next actions into a compact resumable state packet./recoverChecks health, backup integrity, and rollback readiness before an automated workflow advances.An ask_local tool backed by quantized on-device inference. Small, glance-verifiable tasks stay local; concurrency limits and explicit timeouts keep the machine responsive.
Purpose-built, least-privilege interfaces expose only the reads or writes a workflow actually needs. The authorization boundary lives below the prompt.
Scheduled checks verify service health, backup freshness, and restore evidence. Missing proof blocks a healthy-state claim.
Skill installation passes through source review, permission analysis, static checks, and sandboxed execution before trust is granted.
The routing policy. The local model gets work by default when the answer is verifiable at a glance; the frontier model keeps anything where a subtly wrong answer costs more than a slow one. The policy is written down, versioned, and enforced in the agent's instructions:
| ROUTE TO THE LOCAL MODEL | KEEP ON THE FRONTIER MODEL |
|---|---|
| Regex generation, unit conversions, date math | Multi-file code changes, architecture decisions |
| JSON/YAML transforms, cron expressions, git one-liners | Security-sensitive analysis: credentials, auth, permissions |
| Boilerplate: shell scripts, Dockerfiles, config stubs | Tone-sensitive writing meant for other humans |
| Commit message and branch name drafts | Anything needing tools, context, or institutional knowledge |
| Verifiable-at-a-glance answers | Tasks where subtly wrong is worse than slow |
Hermes is my dedicated Windows 11 lab for governed autonomy: a source-controlled control plane, local and frontier model routing, durable memory, scheduled health and backup checks, and bounded workers. It is active systems research, not a production compliance claim.
The interesting part is not that agents can act. It is that authority is explicit, reviewable, and removable. A failed health check or missing proof blocks the claim instead of becoming optimistic prose.
Cross-session memory is the difference between an assistant and a colleague. Mine is an Obsidian vault of plain Markdown: no plugins, no sync daemons, no database. Every file is greppable, diffable, and readable by a human without any tooling at all.
agent_authored: true in its frontmatter and isn't canon until I've reviewed it.