Selected AI Work

RAG · Semantic search

Strategy Sync AI

MSc coursework · Information Retrieval

An information retrieval system that measures how well an organisation's day-to-day action plans line up with its long-term strategy, and explains where they don't.

Role
AI / IR Engineer
Year
2026
Focus
RAG
Stack
5 technologies
  • Sentence-Transformers
  • ChromaDB
  • RAG
  • OpenAI
  • Streamlit
Live demoCode

  • Top-K

    Semantic retrieval per strategy

  • 3 tiers

    Strong · Medium · Weak alignment

  • Live

    Hosted on Hugging Face Spaces

The problem

Strategic plans are abstract, and action plans are operational. They use different words for the same intent, so keyword matching misses real links, and manual alignment reviews are subjective and slow on large documents.

Overview

Strategy Sync uses semantic similarity instead of keyword overlap. Every strategy and action is embedded, each strategy retrieves its closest actions from a vector database, and the system scores how strongly each strategy is supported.

An optional RAG step uses an LLM to write improvement recommendations, and a deterministic fallback keeps runs reproducible. The results appear in a Streamlit dashboard, available as a live Hugging Face Space, with JSON and CSV export.

Architecture

Pipeline

  1. 01

    Ingest

    JSON / PDF

    Strategic and action plans are parsed, and PDFs are converted to structured JSON.

  2. 02

    Preprocess

    Titles, descriptions, and KPIs become clean sentences for better embeddings.

  3. 03

    Embed

    all-MiniLM-L6-v2

    Sentence-Transformers turns each item into a semantic vector.

  4. 04

    Retrieve

    ChromaDB

    A persistent vector store returns the top-K most similar actions for each strategy.

  5. 05

    Score

    Top-3 average similarity per strategy, labelled Strong (≥ 0.75), Medium (≥ 0.55), or Weak, plus overall score and coverage.

  6. 06

    Recommend

    RAG

    LLM-backed or deterministic suggestions for weakly supported strategies.

  7. 07

    Visualise

    Streamlit

    Dashboard, strategy–action graph, and downloadable reports.

Key features

  • Overall alignment score from 0 to 100, with coverage: strategies backed by at least two strong actions
  • Per-strategy breakdown that points out weak or unsupported goals
  • Evaluation against ground truth with Precision@K, Recall@K, MAP, and NDCG
  • Ontology documentation (TTL) ready to extend into a knowledge graph