AI-Enabled Enterprise Knowledge & Agentic Automation Assistant
Python workflows for source-grounded research, structured output and human review.
- Context
- Academic/personal Python workflow and public companion demo
- My role
- Workflow construction, reusable prompts, validation and review steps
- Status
- Prototype workflow; reproducible retrieval companion available below
- Deliverable
- Structured research outputs and a runnable, source-linked example
Decision & evidence
Make the source traceable before acting
I separated information retrieval, synthesis and human review. The public companion makes the retrieval and validation steps inspectable using a fictional knowledge base and extractive output.
The companion uses token-overlap retrieval, not vector embeddings or a live LLM. It demonstrates evidence handling without claiming the original prototype’s model configuration or production performance.
Runnable portfolio companion · fictional knowledge base
A traceable answer, with a review gate
A small offline baseline makes the evidence-handling part of the broader AI workflow visible. These outputs are generated by the downloadable Python script; no live model or external service is used on this page.
- Inputs
- Three fictional JSON documents covering access requests, incident handoffs and change reviews, plus a question.
- Retrieval
- Lowercase token overlap, common-word removal, a minimum of two shared terms and up to two matching documents. No vector database or embeddings.
- Output
- Verbatim supporting extracts, source IDs, a status and a required human-review flag. The original prototype’s LLM synthesis is not reproduced here.
- Validation
- Check source IDs, exact extract text, duplicates, output status and the review gate. Abstain when retrieval finds insufficient evidence.
Inspect a generated example
What should an incident handoff include?
Draft · human review required
An incident handoff records the affected service, observed symptoms, evidence reviewed, current owner and next action. Keep assumptions separate from confirmed observations.
Source: DEMO-HANDOFF · fictional guidance
Review the cited guidance against the question before use.
What must an access request record before fulfillment?
Draft · human review required
Access requests require a business justification and approval before fulfillment. Record the decision and retain completion evidence.
Source: DEMO-ACCESS · fictional guidance
Review the cited guidance against the question before use.
What is the reimbursement limit?
Insufficient evidence · no answer inferred
Request an authoritative source; do not infer the missing policy.
Matching words does not establish semantic relevance or correctness. The human reviewer must assess whether the retrieved guidance answers the question. This example measures no productivity improvement.
Overview
Built a Python-based agentic AI workflow using retrieval-augmented generation concepts, LLM tooling and reusable automation steps for enterprise research and technical documentation.
Objective
Research enterprise information and synthesize complex inputs into grounded technical outputs while keeping validation and human review in the workflow.
My contribution
- Built a Python-based workflow combining RAG concepts, LLM tooling and reusable automation steps.
- Used the workflow to research enterprise information, synthesize complex inputs and generate grounded technical outputs.
- Implemented validation and human-review steps, reusable prompts/workflows and structured outputs.
Technical approach
- Connected research and synthesis steps through reusable prompts and structured output formats.
- Included validation and human review to assess generated outputs before using them for documentation, prototyping or further engineering work.
Scope & considerations
Validation checks the source material and output structure before human review. Unsupported questions should lead to a request for better evidence rather than an invented answer.
Outcome
Demonstrates practical GenAI use for research, documentation, prototyping and engineering productivity, alongside validation and human-in-the-loop automation.