Agents that fit the team, not the other way round

The method

At Princeton Digital Labs I research the method. Four capabilities (product design, software development, software quality, project management) each get their own agent, anchored on the relevant engineering standards, with a testing methodology to assess the agents themselves. The agents are to be released as open source under the MIT license.

The practice

At LSPedia I apply it. The agents are tools for the specialists who use them, so the program is as much a change-management mechanism as an architecture: each agent strengthens a capability the team already has. I built a design system integrated with the assistants, now used by all of LSPedia's business analysts.

In production

I shipped LSPedia's customer-facing AI chatbot, with human-in-the-loop escalation designed in. The model was chosen through an empirical evaluation of GPT-4, Gemini and Claude, and Gemini was deprecated in favour of Claude on reliability and cost. The hallucination-mitigation method behind it (structured prompting, self-consistency checks, source validation, templated decomposition) is now reusable tooling.

Before it was fashionable

At Adents I wrote the business case for an AI supply chain intelligence platform: technology evaluation from graph databases to Azure ML, and machine-learning use cases for vendor evaluation, order fulfilment forecasting and anomaly detection, built with Microsoft. At Princeton Digital Labs I also gave technical guidance to VoiceStream, a commercial multi-agent LLM platform that retrieves, compiles and summarises content and turns it into audio.