Ajanee Igharo

• Built a Python analytics-agent baseline over deterministic synthetic delivery-operations data with dataset-scoped, read-only ClickHouse access, semantic metric guards, and explicit SQL, transport, resource, and dataset boundaries.
• Built deterministic evaluation infrastructure that separates execution success from task correctness, factual consistency, tool grounding, model/tool calls, latency, token use, and repeatability.
• Established a frozen Q1-Q6 benchmark; the accepted single-agent baseline achieved 66.7% task success in each of three runs, with 0.0 percentage-point run-to-run task-success variance, exposing stable failure modes rather than one-off errors.
• Implemented and evaluated an oracle-metadata routed experiment under matched conditions. Runs scored 5/6, 6/6, and 6/6 versus 4/6 in each baseline run; frozen metadata selected deterministic handlers for Q1-Q5, so the experiment does not test natural-language route inference or establish a general multi-agent advantage.
• Maintain automated regression coverage across Python 3.11 and 3.12 with GitHub Actions and CodeQL security analysis.

Technologies: Python · SQL · ClickHouse · Ollama · GitHub Actions · CodeQL · automated evaluation

• Built a deterministic Good Night Check workflow with explicit tool contracts for state inspection, permitted actions, human approval requests, and audit logging.
• Designed risk-sensitive approval gates: low-risk actions can execute automatically, while lock/access uncertainty pauses for human review and failed device calls leave state unchanged.
• Kept device integrations simulated for reproducible testing and wired a Strands Agents adapter while separating live provider behavior from deterministic safety logic.
• Added deterministic tests covering low-risk actions, approval-gated lock behavior, and device-failure paths.

Technologies: Python · Strands Agents SDK · agent tooling · automated testing

• Designed a product analytics case study for fragmented customer-portal journeys, with a measurement framework spanning task completion, follow-through, workflow friction, review cycle time, and closed-loop completion.
• Defined a constrained AI scope for summarizing approved source material, organizing tasks, and drafting questions while prohibiting source-record changes, autonomous decisions, and unsupported high-impact actions.
• Framed the work explicitly as a concept and measurement-design case study based on secondary research, not a production deployment or measured outcome study.

Paul English Applied Artificial Intelligence Institute, University of Massachusetts Boston

• Support applied AI programs, workshops, learning experiences, operational workflows, communications, curriculum materials, and community-facing initiatives across the Institute.
• Co-led planning for a two-day AI and career-development workshop series for UMass Boston career advisors through 10+ collaborative planning meetings, speaker consultations, curriculum development, and rehearsals.
• Led speaker communications, curriculum outlines, and pre-arrival documentation for the series, coordinating materials across student leaders and invited speakers.
• Designed and delivered “AI-Powered Personal Branding,” a hands-on session on evidence-based professional narratives, privacy, accuracy, credibility, and individual ownership when using generative AI.
• Contribute to the Institute’s Student Advisory Council and interdisciplinary AI education, outreach, and program-development work.