Back to work

UBCI · Case 01 · 2026

Banking & Regulatory RAG

Can a banking assistant retrieve the right source first without making every CPU request depend on a slow, heavyweight reranker?

Stacks of indexed paper files representing regulatory document retrieval
Role · Technical ownerSanitized case studyBanking · Information retrieval

The project

The challenge

I designed an end-to-end retrieval and evaluation system for banking content, Central Bank of Tunisia publications, and validated internal documents. The work improved first-result relevance while distilling the expensive ranking behavior into a practical CPU path.

My role

I owned the retrieval architecture, benchmark design, teacher-to-student reranking experiments, API workflow, and the source-grounded user experience.

83.8%BCT Hit@1
82.6%Hit@1 retained
1.55sCPU latency

Approach

Four clear steps

  1. 01Ingest & validate
  2. 02Retrieve twice
  3. 03Fuse & rerank
  4. 04Answer with evidence

Result

What the evaluation showed

Teacher reranking raised BCT Hit@1 by 12.6 percentage points.

Baseline71.2%
Teacher83.8%
Student82.6%

Outcome

What changed

The final workflow joined document validation, hybrid retrieval, reranking, generation, and source inspection in one product path.

Caveat

What this does not prove

Hit@1 measures retrieval relevance, not complete answer correctness or user comprehension.
Next projectNetwork Forecasting & Anomaly Detection