Turning financial data into decisions people act on.

An aggressive lending policy earned £298 more profit per application than a conservative one, despite defaults 3.3 points higher (16.05% vs 12.79%).

• Simulated both policies on LendingClub loan data and tested the gap with a Welch t-test (p < 0.001). Segmented borrowers with K-Means on 11 features (k = 3); the aggressive policy won in every cluster, by £521.60 per application in the highest-risk one.
• Built a neural network that predicts defaults at ROC-AUC 0.68 on 8,417 unseen loans, where a cluster-average baseline has no discrimination within a segment (AUC 0.50).
• Recommended a segmented, hybrid policy, with reject inference and a live pilot before rollout. Results are correlational because the data holds approved loans only.

Logistic regression identified 77% of likely business-credit applicants (AUC 0.854) from 100,000 customers and 200 anonymised features.

• Benchmarked six classifiers in a CRISP-DM workflow, tested two preprocessing pipelines (with and without one feature that was 40% missing) and handled a 10/90 class imbalance with class weights instead of synthetic data.
• Chose logistic regression over XGBoost (AUC 0.861) because XGBoost found only half of applicants (recall 0.50 vs 0.77) and missed applicants mean lost revenue. It is also easier to explain to regulators. Random Forest scored 90% accuracy but found no applicants.
• Flagged 1,549 of 2,008 actual applicants on the test set. Proposed a relationship-manager dashboard pilot before wider rollout.

Reallocating a 300-generator fleet across 1, 4, 8 and 16-week rentals lifts modelled revenue by £466,430 (10%) with no price changes.

• Built a linear programme in Python that raises load factor from 73.4% to 79.1% and ROI from 5.17 to 5.69, cutting idle-capacity weeks from 39 to 11 of 52.
• Proposed using its shadow prices (up to £672.70 per generator in week 37) as booking bid prices. A 16-week rental starting in week 24 earns £5,734 against £4,780 of opportunity cost, so it would be accepted.
• Limits: one branch, a model-based projection, and no customer data to confirm the price-sensitivity assumptions.

Implemented a 14-table MySQL database and generated the synthetic dataset behind a team analysis of £967M AUM, up from £452M.

• Built the schema with primary and foreign keys, CHECK and UNIQUE constraints and a rule linking client status to closure date, normalised to 3NF.
• Wrote the generation specifications for 1,000 clients, 25 advisors and a year of daily account snapshots, with asset prices simulated using geometric Brownian motion.
• The team's analysis found every advisor with a £100M target missed it, and that 2025 added 46 net clients but lost £572K of AUM through churn. Synthetic data, so the findings are illustrative.

Hybrid retrieval, reranking and query rewriting raised retrieval relevance from 0.70 to 0.84 across ten World Development Reports (2016 to 2025). LLM-judge scores for the answers did not improve.

• Indexed 4,589 chunks from 2,854 pages. Fused FAISS and BM25 results with Reciprocal Rank Fusion, reranked the top 20 with a cross-encoder to a top 5 and rewrote queries with HyDE.
• Tested both pipelines on 10 queries of four types. Relevance improved on four and was unchanged on the rest. For a question about the 2016 report, the baseline returned no 2016 chunks, while the enhanced pipeline returned four of five.
• Judge scores (1 to 5): correctness 4.5 for both, completeness 4.9 vs 4.8, grounding 3.8 vs 3.1. The judge saw only a truncated excerpt of the context, so I treat the grounding result as unreliable and would rerun it with the full context.

Relevance is a keyword-overlap heuristic.

An integer-programming model cut the strength gap between World Cup groups to 87 FIFA points, against 500 in the confirmed groups of the official draw.

• Adapted the Ghoniem et al. (2017) model from 32 to 48 teams in Pyomo: 12 balanced groups under a max-min objective, with confederation caps and three co-hosts in Pot 1, solved to a 0.8% optimality gap.
• Assigned groups to letters and stadiums with two further models and host-pinning rules, sending the most popular fixture set (score 20.0) to the 72,766-seat New York New Jersey stadium.
• Limit: only six of the twelve official groups were confirmed when the comparison was made.

Led a four-person team to first place among nine teams in a live Emirates Group challenge, winning a funded trip to Emirates HQ in Dubai.

• Entries were judged on how practically the concept could be implemented in Emirates' existing app.
• Framed the '360-day gap' between flights for Gen Z Skywards members.
• Proposed Runway to Rewards, a gamified micro-engagement concept built on the Hook Model, with a phased commercial rollout.