AI-assisted BI & Report Generation

AI · Business IntelligenceManufacturing / Operations Reporting

An AI-assisted BI tool for a manufacturing client, built on a retrieval-augmented (RAG) approach so business users can query the company's own data just by asking in plain language, instead of waiting on engineers to write SQL or build a custom report each time. Rather than answering from the model's own knowledge, the system retrieves the real context first: vector / embedding search surfaces the relevant tables, columns and business context for the question, the model turns that into a safe, read-only SQL query run against the company database, and the retrieved records feed straight into report generation, so every answer stays grounded in the company's actual data. The hard part was never just calling an LLM, it was making that retrieval accurate and safe, and the result usable for non-technical staff. I worked across the whole path, from PoC to production, in a role close to tech lead.

RAG pipeline diagram: a plain-language chat question flows through vector retrieval, schema and table detection, safety guardrails and the company database, then out to reports and dashboards — How it works: vector search finds the relevant context, safe SQL fetches the company's own data, and the result becomes a report

What I did

Built the RAG retrieval pipeline: vector / embedding search to surface the relevant tables and columns, schema-aware prompting, and safe SQL generation over the company's data
Designed the safety guardrails: SELECT-only queries, LIMIT control, parameterized handling and execution checks before anything runs
Connected AI-generated results into the report-generation flow, so answers came back as usable reports, not a raw table dump
Redesigned the reporting UI so non-technical business users could actually use it day to day
Built across the full stack: the React / TypeScript UI and the FastAPI / Python service behind it

Outcome

Cut routine report creation from days to minutes: business users could answer their own data questions instead of queuing a request with engineering, while every generated query stayed safely constrained.

Built with

ReactTypeScriptFastAPIPythonOpenAI / LLMEmbeddings / vector searchPostgreSQLAWS

Company and client names are withheld. Roles are described as “worked on” or “contributed to” where ownership was shared.

Have a similar problem to solve?

If something here mirrors a process your team is wrestling with, I'd be glad to talk through what an approach like this could look like for you.

Contact me All work