Work

That constraint is the whole problem. Most conversational AI assumes a GPU, and the languages with the least support are often spoken exactly where there is not one. A model that needs hardware the deployment will never have is not a candidate, however well it scores.

My part was evaluation. I benchmarked 10+ AI and large language models against what the stack would actually have to do, integrated 6 key features that improved chatbot user experience by 25%, and produced the build-versus-buy recommendation for the foundation's conversational AI stack.

The efficiency figures are the proof rather than the point. Model resource consumption came down 18% and operational costs 12%, using optimization techniques drawn from LLaMA 2, Bloom and Falcon — balancing performance against cost and sustainability, which is the trade-off that decides whether the thing can run where it is needed at all.

I identified 80 strategic partners and built a tailored acquisition strategy for each, sequenced by ICP fit and expected value, contributing to 50% customer-base growth. I secured $200,000 in funding from SAGE by turning voice-analysis data and impact metrics into an evidence-based investment case.

Net Promoter Score rose 15 points through UX observation studies, A/B testing and process analysis. Partnership reporting visibility rose 30% with Python dashboards across 35 client datasets, built in direct partnership with engineering.

The research design is the part I would carry anywhere. I tested in pairs — one person as user, one as listener, simultaneously — across successive beta releases, which surfaces a class of failure solo testing cannot reach: one side's experience breaks while the other side sees nothing wrong and has no way to know. For a peer-support product that is not a technical fault. It is a person reaching out for support and receiving silence.