Internal knowledge. One assistant.
An internal assistant and the knowledge pipeline behind it: document processing, contextual answers and retrieval evaluation.
- 01Sources
- 02Extract & normalize
- 03Chunks & embeddings
- 04Retrieve context
- 05Grounded response
- Role
- Development & retrieval evaluation
- Product
- Internal tools / AgeVolt
- Period
- 2026
Overview
Alongside the portal, I worked on an internal AI assistant and a pipeline that turns company information into searchable knowledge. My work covered source processing, parts of the assistant interface and runtime, and tools for evaluating retrieval quality.
My work
- Worked on source discovery, document extraction, normalization, chunking and embeddings, with artifacts stored in S3.
- Contributed to the assistant interface and runtime, connecting relevant knowledge with OpenAI Responses API.
- Created and used evaluation datasets and repeatable runs for comparing retrieval configurations.
- Diagnosed index inconsistencies, incremental updates and lexical/vector retrieval experiments.
What I worked on
From source systems to knowledge artifacts
Worked with sources including ClickUp, Microsoft Graph and company documents. The pipeline separates extraction, normalized documents, chunks and indexes, which makes missing content and inconsistent artifacts easier to diagnose.
Assistant interface & runtime
Contributed to the chat workspace, conversation history, feedback and the connection between retrieval and generated answers. Relevant context from the knowledge layer supports responses through OpenAI APIs.
Retrieval evaluation & experiments
Worked on evaluation questions, configuration comparisons and tuning runs, including lexical and vector retrieval experiments. Implementation and validation are described here separately from production rollout; not every experimental extension was deployed.
Result
Implemented and debugged parts of the pipeline and assistant, alongside evaluation tooling for retrieval comparisons. As an internal product, it is presented through its technical flow and my contribution, without private source material or internal screenshots.
Technologies
- TypeScript
- Node.js
- Next.js
- OpenAI API
- AWS S3
- Microsoft Graph
- MySQL
- Retrieval evaluation