services, rates, and how this actually runs.
no discovery-call theater. here's what i do, what it costs, and how fast i respond — so you can decide before we ever talk.
Browser and desktop agents that do real work inside guardrails you control — not a demo that falls over the first time it hits something unexpected.
- Tool-scoped agents with least-privilege access by default
- Persistent memory so agents stop repeating solved mistakes
- Clamped spend and outbound network access, fail-closed
Local LLM inference stacks sized to real hardware budgets — off the per-token treadmill, with the same reliability expectations as a production service.
- Model and quantization selection matched to your hardware, not marketing specs
- llama.cpp / Vulkan / IPEX-LLM stacks, tuned and benchmarked, not just installed
- Real cost math before you buy hardware, not after
Security review for AI systems and the infrastructure around them — from someone who's carried an incident to resolution, not just read about one.
- Agent tool-access and permission audits
- Prompt-injection and hostile-input review
- General infrastructure and MSP-style security hardening
rates & engagement
For scoped work, audits, or ongoing engagements billed as time is spent. Standard for most AI agent and infra work.
For well-defined deliverables — an agent build, an infra migration, a security review with a report. Quoted after a short scoping conversation.
For teams that want ongoing access — a set number of hours a month, priority response, and continuity across engagements.
how an engagement actually starts
Send it through the contact form, email, or Upwork — whatever's easiest. No form-filling ritual required, just the actual problem.
I respond within one business day. If it's a quick question, you get an answer. If it needs real scoping, that's a short written back-and-forth, not a 45-minute call to sell you something.
Fixed-bid work gets a quote before anything starts. Hourly work starts when you say go, and you always know what's been billed.
Progress updates as the work happens, not a surprise at the end. If something in scope turns out to be bigger than expected, you hear about it before it's billed, not after.