Applied agents, workflow accelerators, decision-support tools.
The same systems-thinking from the case studies, applied to agentic workflows. Each system has a clear use case, a defensible reason the agent framing is appropriate, named tools, and explicit human review points — not vague futuristic assistants.
DataPilot
Turns natural-language questions over a lending portfolio into planned SQL, executed queries, and inline visualizations — compressing the analyst round-trip from half-an-hour to seconds.
Why agentic · Multi-step analytical questions require plan → query → interpret → re-query loops. An agent with bounded SQL tool access closes that loop without handing the analyst a wall of broken queries.
ClinicalPilot
Grounds clinical reasoning over de-identified ICU EHR data in the underlying source tables and patient IDs — with guardrails that keep the system out of prescriptive territory.
Why agentic · Clinical questions traverse labs, diagnoses, medications, and cohorts in sequence. A tool-using agent can walk that graph while keeping every claim traceable to a specific record.
MarketPulse
Stitches together Redfin and FRED data across 96 Bay Area zip codes to answer affordability, comparison, and trend questions an analyst would otherwise assemble by hand.
Why agentic · Home-buying questions rarely decompose into one query. The agent plans across six domain tools and returns a structured answer with the comparisons and caveats an analyst would produce.
LifeContext
A concept system exploring how recommendation could feel if it reasoned over time-of-day, mood, and lifestyle context — not just watch history. Treated as a product-design case study in signal selection and tradeoffs.
Why agentic · Re-ranking under context signals is a structured decision problem, not a chat loop. The exercise is about choosing the right signals, naming the tradeoffs, and defending the eval — the kind of thinking a personalization team does before any model is trained.