Sumaiya Shrabony
I help data and operations teams make AI usable at work: not hype, not theory, the operator version that survives real users, governance, and broken dashboards.
Denver, Colorado, United States
Actions
Sumaiya Shrabony is a Technical Program Manager and BI practitioner at the University of Colorado Denver, where she manages enterprise analytics infrastructure across 200+ Azure Data Factory pipelines, SSAS Tabular models, Power BI, 40+ departments, and 530+ users.
Her work sits in the gap most AI talks skip: what happens after a tool looks good in a demo and has to survive real workflows, governance constraints, anxious stakeholders, broken metrics, and users who did not sign up for AI. She writes Ground Truth, a weekly newsletter that tests AI tools against practical enterprise data work and reports the adoption barriers stopping them from landing.
Sumaiya speaks on three subjects: AI adoption translation for non-technical teams, BI and semantic-layer reliability, and agentic workflows that refuse to guess when the user cannot verify the answer. She moved to the US from Bangladesh at 19 and built her technical career without inherited defaults, which now shapes how she teaches: specific, direct, and built for people who need the work to make sense by Monday morning.
Area of Expertise
Topics
Build a 3-Agent System that Refused to Guess: Hands-on Multi-Agent Orchestration with Evals
Most multi-agent tutorials teach you how to make agents talk to each other. This workshop teaches you how to make agents refuse to talk when they shouldn't.
We'll build a working 3-agent pipeline from scratch during the session. Each agent has a single job, a structured JSON contract defining its input and output, and a refusal threshold: a confidence score below which the agent stops processing and surfaces uncertainty to the user instead of passing a bad answer downstream.
What you'll build in 75 minutes:
Agent 1 (Parser): ingests a raw document and extracts structured fields. If any field falls below confidence threshold, it flags the gap instead of guessing.
Agent 2 (Analyzer): takes the parsed output and performs comparison logic. Refuses to run if Agent 1 flagged incomplete data.
Agent 3 (Generator): produces a user-facing output (a summary, recommendation, or draft). Carries a confidence surface that tells the end user what the system is sure about and what it isn't.
The architecture pattern is the one I used for AidLens, a financial aid decoder I shipped in May 2026 for first-generation college students where confident wrong answers cost people real money. But the pattern is framework-agnostic: you'll use it for any domain where your agent pipeline serves users who can't verify the output themselves.
Tech stack for the workshop: Python, OpenAI or Anthropic API (bring your own key or use the shared sandbox), no framework dependency (we'll build the orchestrator from scratch so you understand every decision). You'll leave with a running 3-agent repo on your machine.
What makes this different from a LangChain/CrewAI tutorial:
We design the eval FIRST, then build the agents to pass it (not build first, eval later)
Every agent has a refusal mode, not just a happy path
The orchestrator is 50 lines of Python, not a framework. You'll understand what's happening.
Prerequisites: Comfortable reading Python. Familiarity with LLM API calls (any provider). Laptop with Python 3.10+ and an API key (OpenAI or Anthropic).
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top