Experience · Donald B. Havery

Intelligence Manager and Owner, 2023 to present.

What I own: The AI budget, the model choices, the builds and the results for my businesses, including a live site with Stripe checkout and a production customer-service agent.
Spend and efficiency: Hard cost limits in production: an output-token cap, a chat-history cap, an 8-step limit per run, per-visitor and daily call limits, and a switch that keeps the agent off until it is turned on. The cheapest path that passes its tests: a flat-rate subscription before metered calls, local open models before frontier ones, and a stored answer with no model call when retrieval confidence is high.
Agent, code or person: Each workflow is mapped with the people who run it before anything is built. Every agent hands off to a person with a reason code and the transcript, and a correct handoff scores as a pass in evaluation.
Measurement and governance: A 103-scenario live evaluation with an answer-rate floor that is never lowered without a written reason, a CI gate that fails the build on regression, and compliance tools in which a model may add a note but cannot change a verdict.

Spend efficiency: 58 days of daily agent work, 164,039 model calls, on one flat-rate plan that cost about $380. 96.5% of input came from the prompt cache.Claude Code usage records
Quality and routing: On a 71-question test of the live customer-service agent, wrong answers fell from 3 to 0. Of the 23 cases still handed to a person, 21 were policy the business had not decided, so they were listed as open decisions for the business.
Model cost: A grading step moved from a premium model to a lower-cost one only after the cheaper model passed a calibration set in which about half the answers are built to fail.
Capacity: Parallel agents cut from 20 to 5 after a burst used a five-hour plan allowance in about 90 minutes, with a per-agent GPU limit so several workloads share one machine.

Took a live agent offline: When the customer-service agent answered questions the business had not decided, I turned the agent off and left the site and payments running. It came back on only after each new guard was proven by a test that fails without it.
Sent policy gaps to the owner: Most of the remaining handoffs were business decisions, not wording problems, so they went to the owner to decide and write down instead of to the prompt.
Flat rate before metered: My build work runs on a flat-rate plan. Metered API calls are used only inside products, behind caps and a kill switch.
Quality before cost: Local models are the default, but a frontier model is used where the local one fails the quality bar, as in the text-to-CAD tool, where small models place multi-feature geometry wrong.
Checked model licenses before publishing: One generation model’s license excludes the United States and covers its outputs, so its output stays on my machine and out of every published or client-facing product.
Audited my public claims: Every public claim is checked against the code. When the gateway’s cache was described as semantic and the code matched requests exactly, the description was changed to exact-match.
Priced to the market: Set the studio’s prices at the low end of the market range and cut the discount stack from 22% to 12% so it could not quote below credible providers.

Web and AI studio: A live studio site that quotes from one price book, takes deposits through Stripe, and runs the production customer-service agent. Case study.
Grooming brand, 2017 to 2021: Direct to consumer: formulation, small-batch manufacturing, packaging, marketing, fulfillment and customer service, with full profit-and-loss responsibility.
Fuel-delivery operations platform: Designed the franchise economics (royalty, brand fee, per-vehicle technology lease) and a nine-role operating platform with a public demo. Pre-launch.
Investigations practice: Research, surveillance, case management, evidence handling and client reporting.
CDL training site: A course in which every lesson is tied to the regulation it comes from. truckingtutor.com.

Old Dominion Freight Line, 2022 to present: Line-haul driver at Old Dominion Freight Line. Class A CDL with hazmat, tanker and triples endorsements, in an operation run on dispatch and federal hours-of-service limits.
Peninsula Truck Lines, 2017 to 2021: Line driver and company-certified driver trainer: assessed skills, gave road instruction, and signed off new operators to drive unsupervised.
Alpha Transport, 2012 to 2017: Fuel tanker driver: hazmat, tanker and oversized loads across Western US routes under FMCSA rules.
Supervisor roles: Supervised people at several employers and in my own businesses.

← Back to portfolio Résumé Contact