Service 03 / AI Integrations
Practical AI, in production.
I build AI into systems that already do real work — customer support, document handling, retrieval over your own records, content workflows — for businesses in Claremont and the Pomona Valley. Scoped around a specific job with a measurable outcome, with the running costs sized before the build and a review step wherever the output has consequences.
- Scoped around
- One job
- Cost modeled
- Before build
- Based in
- Claremont
01 — What you get
What ships.
One job, done properly
A defined task with a measurable before and after — hours saved, response time, error rate — rather than a general assistant nobody is quite sure how to evaluate.
Retrieval over your own material
Answers grounded in your documents, records, and history, with the source attached. A model that cites where an answer came from is one a person can check.
A human in the loop where it counts
Review steps, confidence thresholds, and constrained outputs on anything that touches a customer or a record. Drafting is safe to automate; deciding usually is not.
Cost and failure modeled up front
Per-request cost estimated before building, caching where work repeats, spend limits enforced in code, and a defined fallback for when the provider is down or slow.
02 — How it runs
Prove it small, then widen.
Step 1
Find the job worth doing
We look at where time actually goes and pick the task with a clear success measure. Some of these turn out not to need AI, and that is a good outcome to reach in week one rather than month three.
Step 2
Build a narrow version
The smallest thing that does the job on your real data, evaluated against examples where you already know the right answer. Cheap to change while the approach is still in question.
Step 3
Harden and integrate
Error handling, rate limits, spend caps, monitoring, and the review workflow. Then it goes where the work actually happens instead of living in a separate tool.
03 — Proof
Evidence-led, not generically encouraging.
Eeva Learning analyses spoken English for advanced Korean speakers and returns the specific expressions and patterns worth refining. The hard part was never generating feedback — it was making feedback that starts from the English the learner already used, so the response is evidence a person can act on rather than praise they cannot verify. That is the same discipline every one of these builds needs.
See the work04 — Where I work
Based in Claremont. Working across the Pomona Valley.
AI projects go wrong when the person building has not seen the work being automated. Being in Claremont means clients across Pomona, Ontario, and Rancho Cucamonga can show me the actual inbox, the actual paperwork, the actual bottleneck — which is usually where the real scope is hiding.
- Claremont, CA
- Pomona, CA
- La Verne, CA
- San Dimas, CA
- Upland, CA
- Montclair, CA
- Ontario, CA
- Rancho Cucamonga, CA
05 — Questions
Before you ask.
Is this just adding a chatbot to my site?
Usually not, and a chatbot is often the wrong answer. Most businesses get more from AI doing one specific job well — reading incoming documents, drafting a first response, routing a request to the right person, searching your own records in plain language — than from a general assistant that talks about everything and is trusted with nothing.
What does it cost to run once it is live?
Model usage is billed per request, so the running cost scales with how much you use it. I size that before building and design around it — caching repeated work, using smaller models where they are sufficient, and putting hard limits in place so a bug cannot produce a surprise invoice.
What about our data — does it go somewhere we cannot see?
That is a decision we make explicitly, not a default you inherit. Which data leaves your systems, which provider receives it, what their retention terms are, and what stays local are all part of scoping. If the honest answer is that the data cannot leave, the design changes to match.
How do you know the output is actually right?
By not trusting it blindly. Anything with consequences gets a review step, a confidence threshold, or a constrained output format the system can validate. AI drafts and a person approves, or the model picks from a fixed set of options rather than generating free text into your database.
What if AI turns out to be the wrong tool for this?
Then I say so and we build the boring version. A well-written rule, a search index, or a better form eliminates more manual work than a model does, for a fraction of the cost and none of the uncertainty. The goal is the workflow being fixed, not AI being used.
Have an idea you want built properly?
Describe the task you would hand to a new employee if you could — the repetitive one that eats an afternoon a week. That is usually the right first project, and it is a far better starting point than deciding to use AI and looking for somewhere to put it.