Brian Wolfe

AI consulting. I design and build LLM systems, autonomous agents, and retrieval that hold up in production — with the guardrails and evals that keep them safe to run.

Get in touch

What I do

AI agents & automation

Agents that take real actions on real systems — with human-in-the-loop approval, fail-closed guardrails, and tamper-evident audit trails, not just a clever prompt.

LLM & MCP integration

Wiring models and tools into your stack: multi-model routing by task, cost, and privacy; MCP clients and servers; tool use that degrades gracefully when a provider is down.

Retrieval & memory

RAG built properly — Postgres + pgvector, hybrid full-text and vector search with rank fusion, reranking tuned to your data. Not a demo that falls over on a real corpus.

Private & self-hosted AI

For teams whose data is the business: local-first and on-prem model runtimes, with hard guarantees — enforced by tests — about what never leaves your infrastructure.

Selected work

Self-directed builds over the past year — each shipped end to end, solo, and backed by real test suites.

Private conversational AI

A self-hosted AI engine with durable memory

A conversational AI that remembers, runs on your own hardware, and treats privacy as a hard constraint rather than a setting.

  • Multi-model router that chooses per turn between local open-weight models and cloud APIs by task, cost, and privacy.
  • Long-term memory on Postgres + pgvector — hybrid full-text and vector retrieval with reciprocal-rank fusion and importance/recency reranking.
  • Privacy enforced by tests: a suite that guarantees sensitive text never reaches a cloud provider — a guarantee in code, not in a policy doc.
  • ~2,400-test regression gate; an optional multi-tenant layer with auth, metering, payments, and fail-closed safety gates.

Computer-vision automation

An Android automation platform & visual workbench

A platform that drives real Android devices from what's on screen, plus a no-code workbench for building and sharing those automations.

  • Vision-driven control: OpenCV template matching and cross-platform OCR (EasyOCR on Windows, Apple Vision on macOS) driving devices over ADB.
  • An automation IR — one serializable step-tree that a recorder, a drag-and-drop block editor, and a code view all round-trip through.
  • Multi-tenant cloud control plane: a websocket relay, accounts/quotas/billing, and on-demand cloud Android emulators streamed to the browser.
  • Over 100k lines across the platform; thousands of automated tests; CI across a Python version matrix and a signed Windows installer pipeline.

Real-time control systems

A solar-aware EV & home-battery charge controller

A deployed controller that diverts surplus home solar into an EV instead of exporting it at poor utility rates — while guaranteeing the home batteries still reach target charge by sunset.

  • Physics-based solar forecasting calibrated against real meter data, feeding a 72-hour battery trajectory model with a hard floor it will never cross.
  • Safety-layered device control: dry-run by default, advisory mode, rate-limit and stale-data guards, geofencing, and a persisted decision audit trail.
  • Deployed via Docker on a cloud host; 100+ tests including a historical backtest simulator that replays past days.

Safe autonomous agents

An AI operations agent for a small e-commerce business

An agent that runs the data-driven operations of a live online shop: it computes true net profit, makes low-risk improvements on its own, and routes anything touching money to a human.

  • Five-tier fail-closed autonomy: unknown actions default to "requires a human," and destructive or terms-of-service-risky actions are hard-blocked, never auto-run.
  • Hash-chained audit ledger: every action is SHA-256-linked to the one before it and independently verifiable — if the system can't prove what it did, it doesn't ship.
  • Live in production against real marketplace APIs, with a research-to-decision paper trail behind every rule.

MCP & local-first data

A personal-finance automation engine on MCP

A local-first dashboard and automation engine built as an MCP client against a live financial data source — structurally unable to move money by design.

  • Rules-based insights engine (cash runway, duplicate charges, subscription audit, and more) that re-runs on every sync and remembers what you've dismissed.
  • Diffed automation ledger that classifies each item as new, continuing, snoozed, or resolved and renders a daily brief.
  • Read-only data scope, localhost-bound, CSRF-guarded — safe by construction, not by trust.

How I work

Hands-on
I write the code and ship working software, then leave your team able to run and extend it. Engagements range from a focused architecture review to building a system end to end.
Evals first
Every system ships with the tests that prove it works and keep it working — including the safety properties that must never regress.
AI-native delivery
I run fleets of AI coding agents under real engineering discipline — isolated workstreams, verification gates, nightly audit routines — so a solo practice ships at a pace that used to take a team.

About

I'm an airline pilot who builds production software. Over the past year I've designed and shipped a range of AI systems solo — from a privacy-first conversational engine with durable memory to autonomous agents that operate real businesses under strict guardrails.

Aviation trains you to think in checklists, failure modes, and the things that must not go wrong. I bring that same discipline to AI systems, which are too often built without it — fail-closed defaults, audit trails, and evals that make "it works" something you can prove instead of hope.

Contact

The best first step is an email describing what you're trying to build or fix: brian@brianwolfe.ai.