work

Problem:Agentic workflows drift into over-engineered plans and harmful actions because nothing interrupts pattern inertia mid-run.

Approach:Wrapped Anthropic's Model Context Protocol in a CPI oversight server that interrogates every tool call, adding reflection and rollbacks before execution.

Outcome:Across 153 evaluation runs CPI doubled task success (27% → 54%) and almost halved harmful behavior (83% → 42%).

Problem:Student collectives lose operational memory each semester because academic assets live in ad-hoc folders with zero conventions.

Approach:Encoded Git-style naming, templates, and checklists directly inside Google Drive so contributors can drop-in without touching code.

Outcome:Synapse biotech cohorts inherit organized archives, alumni logs, and lab manuals without onboarding bottlenecks.

Problem:Founder-led sales teams lose qualified prospects because follow-ups are slow, generic, and emotionally tone-deaf.

Approach:Orchestrated persona modeling, psychological tone mapping, and GPT-4 agents to draft and score outreach grounded in each lead's funnel state.

Outcome:Pilot campaigns recorded 56% higher intake engagement, 45.9% better sentiment, and 17% more qualified leads than templated messaging.

Problem:Scaling frontier AI by brute force is thermodynamically and financially unsustainable for labs and alignment teams.

Approach:Defined twin metrics (RSRC_t and RSRC_i) plus feedback loops that score both training and inference efficiency for any agentic system.

Outcome:Published through MURST with an archival DOI so teams can benchmark throughput trade-offs without runaway spend.