Challenge
Support questions often need both semantic retrieval from policy documents and deterministic reasoning over structured order data. A language model alone should not decide business eligibility.
Selected work
Prototype · being rebuilt
Agentic support intelligence
An experimental retrieval workflow combining order histories, policy documents and task-specific tools for customer-support questions.
My role
I prototyped document and CSV retrieval, custom tools and a multi-step agent workflow using open language models.
Outcome
The prototype proved the integration path but also exposed rule and timestamp assumptions that need to be moved out of the agent before this can become a reliable system.
At a glance
2data modalities
RAGgrounding pattern
WIPcurrent status
System view
The workflow.
A simplified view of the stages and boundaries that shape the project.
- 01Question
- 02Intent routing
- 03Policy retrieval
- 04Order lookup
- 05Rule engine
- 06Grounded answer
Approach
Decisions that shaped the work.
Separate retrieval from decisions
The rebuild will keep semantic policy retrieval in the language layer while moving eligibility rules into deterministic functions.
Use structured history
Order events need explicit types and chronological guarantees rather than relying on ranking order from text retrieval.
Evaluate before adding autonomy
A labelled set will test retrieval, tool selection, rule accuracy and answer grounding independently.
Findings
What the evidence says.
- The current notebook is a prototype and is not presented as production-ready.
- The most important next step is correctness evaluation, not adding more agent steps.
- The rebuild will include tests, a small API and traceable failure cases.
Engineering reflection
This project remains visible as work in progress because showing the engineering gap between a demo and a dependable AI product is part of the story.