01 / SELECTED WORK

The work, in context.

Real problems. Different methods.
A practical reason for every model.

01 — APPLIED AI IN PRODUCTION

A recovery call.
A clearer next step.

Turning call recordings into summaries, case details and recommended actions, so recovery agents can get to the information they need.

CONTEXT: Call Assist Ltd · Insurance & recovery
MY WORK: End-to-end GenAI pipeline, API delivery and deployment

Python / Azure OpenAI / FastAPI / MLflow

Read the project notes

The problem

Manual review meant working through call recordings to find the information needed for a case. The pipeline brings transcription, summarisation, intent and sentiment extraction into one workflow.

What I built

Speech-to-text feeds an LLM pipeline that extracts key case details and generates recovery options and recommended actions. FastAPI serves the outputs, and a Microsoft Copilot chatbot makes the insights available within agent workflows.

Putting it into production

Deployed on Azure, with MLflow for experiment tracking and model versioning, and GitHub Actions for testing and deployment.

Proprietary work. The diagram describes the workflow; it contains no customer recordings or internal system data.

CALL INTELLIGENCESYSTEM OVERVIEW

Speech-to-text turns call recordings into text for the next stage of analysis.

Illustrative architecture · Based on deployed work

02 — ADVANCED RAG APPPERSONAL PROJECT

Ask your documents.
Inspect the evidence.

A document question-answering app that makes retrieval visible. Ask about PDFs, slide decks or notes, then inspect the source passages, reranker scores and retrieval stages behind the answer.

RETRIEVAL: Dense search + BM25, rank fusion and cross-encoder reranking
EVALUATION: Stage-by-stage recall, MRR and answer / abstention checks

Python / LangChain / Chroma / BM25 / Sentence Transformers / Azure OpenAI

Explore the repository

Read the engineering notes

Two ways to find the evidence

Dense retrieval captures semantic similarity; BM25 catches exact terms. Reciprocal Rank Fusion combines the rankings, and a cross-encoder reranks the pooled passages against the question before Azure OpenAI generates an answer from the selected context.

An index that remembers what changed

A content-hashed file manifest and stable chunk IDs allow incremental indexing: changed files are reprocessed, deleted files are removed, and unchanged documents stay indexed.

Measure each stage

The evaluation harness compares dense, sparse, fused and reranked retrieval against a golden set. Saved baselines make changes to chunking and retrieval measurable; answer checks also include questions the documents cannot answer.

In the checked-in reference run, reranking improved recall over fusion, while BM25 alone outperformed dense retrieval. Those results guide the next iteration rather than becoming a general accuracy claim.

Personal project with a terminal interface. Reranker scores are visible but do not yet filter out weak matches; abstention is instructed in the answer prompt. Evaluation results apply to the reference corpus.

FOLLOW THE EVIDENCE02 / RAG

YOUR QUESTION + INDEXED DOCUMENTS

A / SEMANTICDense searchChroma + embeddings

B / EXACT TERMSBM25Keyword retrieval

COMBINEReciprocal Rank Fusion

RERANKQuestion × passageCross-encoder relevance scores

An answer with its sourcesInspect passages · Compare retrieval stages

Conceptual workflow · Source code and evaluation in the repository

03 — AGENTIC AIPERSONAL PROJECT

Verify the claim.
Keep people in control.

AI Insurance Claim Verification Agent

A multi-agent system that ingests a claim, checks it against policy documents and flags inconsistencies. Deterministic checks and human review gates govern the outcome.

AGENTS: Document parsing, policy verification and fraud triage
REVIEW: Low-confidence and high-value claims go to a human reviewer

LangGraph / Azure OpenAI (direct API) / FastAPI / Python / Pydantic / Docker

Read the system design

A defined graph of responsibilities

LangGraph coordinates specialised agents through explicit state transitions. Document parsing, policy verification and fraud triage each have a defined role in the claim verification workflow.

Rules govern; language models assist

The decision core uses deterministic, auditable rules. LLMs extract information, compare it with policy evidence and generate explanations. Pydantic validates the structured data passed through the system.

Evidence before a decision

PII is masked before LLM calls. Grounding checks compare generated claims against source policy documents, and the system refuses to decide when evidence is insufficient.

A useful handoff to a reviewer

Confidence-based routing sends uncertain or high-value cases to a review queue with a structured summary of completed checks, unresolved questions and supporting evidence. A claim can only clear through a deterministic check or a human gate.

Personal demonstration project. The diagram describes the design; it shows no real claims or policyholder data.

A CHECK AT EVERY STEP03 / AGENTS

CLAIM + POLICY DOCUMENTSMask personal informationBefore any LLM call

01
Parse documentsExtract & validate structured details
02
Verify the policyCross-check details against source evidence
03
Triage inconsistenciesFlag risk and unresolved questions

DETERMINISTIC CHECKS + CONFIDENCE GATEIs the evidence sufficient?

CHECKS PASSEDTraceable outcome

UNCERTAIN / HIGH VALUEHuman review

Insufficient evidence → withhold the decision.

Conceptual workflow · Personal project

04 — LIFE SCIENCES × MLRESEARCH

MOLECULAR ACTIVITY / CONCEPTUAL WORKFLOW

Finding signals in
molecular structure.

Predicting anticancer activity from molecular descriptors and fingerprints, connecting my pharmacy background with machine learning.

RDKit / scikit-learn / Ensemble methods

Explore the approach

RDKit generates molecular descriptors and fingerprints. Feature selection prepares the inputs for ensemble classifiers, assessed with cross-validation on unseen molecules.

This MSc / life-sciences project explores how structure-based models can help prioritise candidates for further investigation.

Research project. The visual is a schematic, not a molecular result or a validated clinical claim.

05 — CLINICAL NLPCODE ON GITHUB

TEXT → FEATURES → CLASSIFICATION

Structured signal

CONCEPTUAL VIEW / NO PATIENT DATA

Making clinical text
machine-readable.

An NLP pipeline that classifies gout-related clinical text, turning unstructured notes into information a model can work with.

Python / TF-IDF / scikit-learn

Explore the repository

Read the method

Text preprocessing and TF-IDF vectorisation feed a supervised classifier. Hyperparameter search tunes the model; precision, recall and F1 are used to assess classification performance.

A clinical-text classification project, not a diagnostic tool. No patient records are shown here.

06 — FORECASTING & OPERATIONS

Planning ahead
of demand.

Time-series forecasting on pharmaceutical e-commerce data to support inventory planning and reduce stockouts. Clean SQL data in; planning insights through Power BI out.

Python / Time series / SQL / Power BI

Read the project notes

Transactional and inventory data is cleaned in SQL, then used to build demand forecasts by product category. Results are surfaced in Power BI to make planning easier for operational teams.

Applied at Inletware, alongside dashboards for sales, inventory and fulfilment.

PAST DEMANDBETTER PLANNING