← All selected projects← Back to homepage
In progress2026 — present

On-Device RAG Copilot for KYC/AML

An in-progress, Microsoft-linked MSc research project building a local, evidence-grounded policy copilot for KYC/AML analysts with cited answers and explicit human-review boundaries.

Role

  • Independent researcher
  • AI/ML engineer
  • Microsoft-linked MSc Independent Research Project

Stack

  • Python
  • Microsoft Foundry Local
  • Local SLM
  • Hybrid RAG
  • Sentence Transformers
  • BM25
  • Streamlit
  • Pytest

Problem

Compliance analysts need to locate the governing policy evidence behind an onboarding question without delegating the final decision to an opaque model.

Constraints

  • The pipeline must run locally and retain an offline path that requires no model download.
  • Synthetic case facts must remain separate from the normative policy evidence used for retrieval.
  • Answers must cite retrieved material, abstain when evidence is insufficient and remain decision support for human review.

My contribution

  • Building the policy ingestion, section-aware chunking, local indexing, hybrid retrieval and cited-answer pipeline.
  • Integrating Microsoft Foundry Local while maintaining deterministic hashing and mock-model fallbacks for offline development and testing.
  • Designing an evaluation harness that compares RAG, keyword and no-retrieval modes across retrieval, citation, abstention, task-success and latency measures.

Architecture / methodology

  • Index policy rules as normative evidence while passing synthetic customer cases separately as query facts.
  • Fuse dense and BM25 ranks, preserve rare entity matches and verify cited chunk identifiers after generation.
  • Expose the workflow through a CLI and Streamlit interface with an explicit human-review disclaimer.

Evaluation

  • Use a labelled KYC/AML benchmark to compare RAG, keyword and no-retrieval baselines.
  • Exercise the offline floor with a deterministic hashing embedder, mock LLM and Pytest suite.
  • Track retrieval hit-rate, evidence recall, citation validity, abstention accuracy, task success and cold/warm latency without presenting target thresholds as achieved results.

Current status

  • The project is in progress; the repository documentation defines the end-to-end local RAG architecture, responsible-AI boundary and evaluation plan.
  • Final benchmark results and completion claims are intentionally withheld until the real embedder and local SLM evaluation is complete.

Evidence and links

  • The project is an Imperial College London MSc Applied Computational Science and Engineering Independent Research Project linked to Microsoft.

    Private IRP README supplied for portfolio content review

  • The documented design separates policy evidence from case facts and includes local generation, citation verification, abstention and comparative evaluation.

    Private IRP README supplied for portfolio content review