AI agents
Agents for intake, triage, research, routing, review, and follow-up with approval gates where the workflow needs accountability.
AI automation servicesSacramento, CA · clients across the US
AI automation
for teams that need production systems.
I build AI agents, RAG assistants and workflow automation for businesses across the US — from Sacramento, where I am based, to a trucking carrier in Kentucky that files its fuel taxes through a system I built. The work is closer to boutique software engineering than template automation: fewer manual steps, clear source grounding, custom integrations, and QA before launch.
- Usually a same-day reply
- A free 30–45 minute call
- A written fixed quote before any work
Capabilities
One automation stack, four kinds of leverage.
Every engagement draws on the same four capabilities: AI agents for repeated decisions, RAG assistants for internal knowledge, integrations that stop manual re-keying, and QA that keeps all of it trustworthy.
RAG assistants
Cited assistants over product docs, SOPs, policies, tickets, PDFs, spreadsheets, and internal knowledge sources.
Workflow integrations
Automation across forms, CRMs, spreadsheets, APIs, email, dashboards, and internal tools so data stops being copied by hand.
QA-tested production AI
Continuous evals, deterministic checks, regression cases, edge-case coverage, and launch gates before automation touches real work.
Approach
Production AI engineering, not template automation.
The work I take on is custom automation where reliability, proprietary systems and data handling matter — and where a wrong answer costs money. That is also where a tested system pays for itself fastest.
Built past the demo
AI systems are tested against edge cases, source-grounding failures, hallucination risk, and regression cases before they are trusted with customer or staff workflows.
Custom systems, not template glue
The work can include non-standard APIs, internal databases, dashboards, web apps, iOS interfaces, and approval flows when off-the-shelf automation cannot cover the process.
Workflows where a wrong answer costs money
Good targets include compliance filings, document intake, reconciliation, reporting, scheduling, logistics operations, client verification, and data-heavy review work.
Exact things stay exact
Calculations live in deterministic code, language and judgment go through a model, and anything irreversible waits for a person. The product around it makes every output reviewable.
Where your AI runs
Built on infrastructure I run myself.
Your AI can run on a private, self-hosted stack I build and operate — free local models by default, paid cloud only when it earns its place. For you that means lower running costs, real data privacy, and no surprise bills.
Predictable AI costs
Every AI call runs through one gateway with automatic failover and a hard spend cap — so an automation can't quietly run up a bill against you.
LiteLLMFailoverBudget cap
Local inference and memory
Dedicated machines run free local AI for chat, code, and search, with a vector index and database holding your project's knowledge — no per-token meter ticking.
OllamaQdrantPostgres
Private and reliable
Everything sits on a private network — nothing public — and every service is monitored, so a problem shows up on my dashboard before it shows up in your workflow.
TailscalePrivate meshMonitored
Private by default — nothing is exposed to the public internet.
Process
Four phases, each one visible.
You see the automation working on your real data before launch, not a slide about it. Every phase ends with something you can open.
Scope and fixed quote
You show me the process as it runs today, with the exceptions. I write down what gets automated, what is explicitly out, what I need from you, a date and a fixed price. Nothing is built before you approve it.
Build in slices
The deterministic core first, then the model, then the integrations — each slice deployed where you can see it, with its tests and evaluation cases growing alongside.
Run alongside the manual process
The automation runs on real work next to the way you do it now until the outputs agree. Disagreements become test cases, not surprises after launch.
Launch and hand over
Monitoring and alerts on, the runbook written, the repository and accounts in your name. Changes afterwards are quoted per change or at $150 an hour, with no retainer.
What you own at the end
- The repository, with the tests and evaluation cases that run on every change
- Every account and credential the system uses, in your name
- A written runbook: what runs where, what the alerts mean, who to call
- A fixed price in writing before any work, and the number does not move unless the scope does
- An honest read at the first call, including when a platform you already pay for would do the job
What this is not
- A no-code tool configured for you — if that genuinely fits, you will hear it, and it will be cheaper
- Model training or fine-tuning as a product; the systems here use existing models well
- Automation that sends, pays, files or deletes without a human approval step in front of it
- An open-ended retainer; scope is agreed, built, tested and finished
Proof
Proof over promises.
Every claim on this page traces back to shipped systems: real clients, real tests, real filings — and code you can read.
IFTA filing agent
In production for a Kentucky trucking carrier: deterministic Python computes the quarterly fuel-tax return, a 17-tool review agent checks it, a person approves the filing. 562 automated tests, including a backtest that matches a real state filing to the penny.
Read the case studyVerdict, an open-source QA agent
The testing discipline as a product: it runs unattended nightly against a production codebase, has found real defects in libraries other people depend on, and signs its own run history so a copied result cannot pass for a real one.
See the projectSales agent across eight marketplaces
Unattended listing and repricing behind one adapter interface, a retry state machine, and a human-approval gate on anything irreversible — so a scheduled run never double-posts and never reprices without sign-off.
Read the case studyPricing
A fixed quote, agreed first.
AI automation from
$8,000
fixed quote · $150/hr advisory
Projects start at $8,000 as a fixed quote; in my quotes, most first projects for a small business land between $8,000 and about $20,000 and take two to six weeks. Advisory work, when you want an engineer's judgment rather than a build, is $150 an hour. You get the number in writing before anything starts, and it does not move unless the scope does.
Discuss an automation- Usually a same-day reply
- A free 30–45 minute call
- A written fixed quote before any work
By area and industry
Where this work happens.
Most projects run remotely, wherever you are. For the areas where on-site time is easy, and for the one industry with a system already in production, there is a page of its own.
San Francisco
Prototype-to-production AI for startups: agents, RAG assistants, evals, and launch QA.
Open pageBay Area
Cross-industry operations automation: intake, reconciliation, reporting, and back-office workflows.
Open pageSacramento
Workflow automation for local operators, reporting, logistics, and compliance.
Open pageSan Jose
Document, support, and compliance automation for hardware, enterprise SaaS, and engineering teams.
Open pageOakland
Freight paperwork, dispatch, and recurring reporting for East Bay operators.
Open pageTrucking & logistics
IFTA filing in production, freight document extraction, dispatch and compliance workflows.
Open pageQuestions
What teams ask before starting.
Service-level answers. The area pages cover what automation looks like locally.
What does AI automation actually mean in practice?
Taking a workflow your team runs by hand — intake, triage, document review, reconciliation, reporting, follow-up — and rebuilding it so software does the repetitive part and a person handles the judgment. In practice that usually means exact calculations stay in deterministic code, language and classification go through a model, and anything irreversible waits for human approval. It is closer to custom software engineering than to configuring an off-the-shelf tool.
How do you decide whether a workflow is worth automating?
Good signals: someone does it every week, it involves documents or structured files, the rules can be written down, and mistakes are costly. Poor signals: it depends entirely on relationships, it happens twice a year, or the rules change every time. If a workflow falls into the second group, you will be told that at the intro call rather than sold a project that cannot pay for itself.
Do you work with businesses outside California?
Yes — most of this work runs remotely, and clients outside California are normal: the trucking carrier running the IFTA agent is in Kentucky. I am based in Sacramento, where on-site time is easiest, and Bay Area projects get on-site time when it helps.
Do you replace the tools we already use?
Usually not. Most of the expensive manual work happens between systems — form to spreadsheet, email to record, invoice to ledger — rather than inside any one of them. Automation connects the tools you already run through their APIs instead of asking your team to learn something new and migrate their data into it.
How do you stop an AI system from making things up?
Three ways, layered. Arithmetic and rule-following stay in deterministic code, so the model never does maths. Retrieval-based answers are grounded in your own documents with citations, and the system is built to say it does not know rather than guess. Then evaluations built from your real cases run on every change, so a regression shows up as a failed test rather than as a wrong answer in front of a customer.
Can you take over an AI prototype our team already built?
Yes, and it is one of the more common engagements. The work starts with a review of the existing code and its failure modes, then a written scope for making it reliable: a service account instead of a browser login, jobs that run unattended, structured outputs, error handling and retries, guardrails, monitoring, evaluations, and regression tests — the gap between something that demos well and something you can run the business on.
What does a project cost, and how long does it take?
AI automation projects start at $8,000 against a fixed quote, with advisory work at $150/hour; in my quotes, most first projects land between $8,000 and about $20,000. One workflow, built, tested and deployed, takes two to six weeks depending on the systems involved. Scope and price are agreed in writing before any build work starts, and there are no open-ended retainers.
Who owns the system once it is built?
You do. All the code, all the accounts, all documented and handed over. There is no licensing arrangement, no per-seat fee, and no dependency on me to keep it running. The point of documenting and testing the handover is that your team can maintain it without calling anyone.
What happens when the automation gets something wrong?
The system is designed around that question rather than assuming it will not happen. Deterministic code handles calculations, review gates sit before anything irreversible, tests and evaluations run before launch, and the automation runs alongside your existing manual process on real work until the outputs agree. When something does fail, it fails visibly — flagged as an exception — instead of silently.
Start with one workflow
Send the manual process you want to reduce.
Describe the workflow, the files or systems involved, and what a successful result would look like. You will get an honest read on whether it is worth automating — including when it is not.