I’m building agents for customer calls and messaging. Answering a question is one part of the job; the system also needs to record the request, perform the allowed action and give staff enough information to take over when needed.

Customer operations

I work directly with frontline staff to understand the requests they handle and the backlog they face. The builds combine voice and SMS with existing tools, explicit action boundaries and a path for staff to take over.

The goals are less manual work, shorter caller waits and fewer frustrating handoffs. This work is ongoing.

Software delivery and continuity

I’ve built harnesses and a ChatGPT plugin that let people and agents resume work across conversations with tasks, decisions and source context intact.

The implementation uses Python/SQLite persistence, API and MCP integrations, a hosted service with OAuth/PKCE, and checks for stale decisions and repeated changes. I direct the implementation and work hands-on with the systems.

Benchmarks, evals and agent coordination

One scoped-retrieval fixture reduced the response from 113,598 to 3,513 bytes while retaining the checked work and evidence. The delivery harness uses versioned state, replay-safe changes and specialist handoffs. When a worker says it has finished, I check its result against the current requirements and evidence.

Technical detail: the benchmark, evaluation layers, feedback loops and five-specialist pilot.

Verification during execution

I check whether the intended action happened, whether it happened once, and whether the result matches what the user was told. I run those checks at the relevant steps in the workflow and test the conversation separately.

More experience and outcomes