What survived the process.

Mishi is a three-role WeChat mini-program for employees, stall operators and administrators across Inspur's internal canteens. During a seven-week internship, I owned the employee-facing module: restaurant and dish discovery, search, multi-stall cart management, pre-orders, pickup codes, order tracking, reviews, likes and favourites.

I rebuilt the ordering pipeline around consistent API fields, vendor-level grouping, preserved price precision, synchronised lifecycle states and duplicate-submission protection. I also added an authenticated Node.js and Express review endpoint with validation and ownership checks, and refactored the interface across 105 files into reusable design tokens and centralised theme configuration.

Open the Mishi repository →

A first-year algorithms project combining implementation, experiment and proof. I wrote bubble sort, merge sort, quicksort and counting sort from scratch, then timed them on inputs from 100 to 10,000 elements and plotted their growth on a logarithmic scale.

The repository also preserves the presentation, speaker notes, visual assets and my reconstruction of the decision-tree lower bound for comparison sorting. The important lesson was learning to distinguish recognising a proof from being able to rebuild it.

Open the Big-O repository →

I led the end-to-end development of a cross-sectional prediction pipeline across nine global automotive equities, contributing roughly 80% of the quantitative architecture and core code. The model uses ranked momentum, RSI and volatility features with strict temporal separation and a shallow Random Forest to look for reversal signals in noisy price series.

The 12-month out-of-sample test produced a daily Spearman Rank IC of 0.0386. I treat that as a research result, not a live-trading claim: turnover control, transaction costs and signal smoothing remain necessary extensions before the model could support an investable strategy.

GitHub repository →

Our cross-university team acted for Samsung Electronics, selected Lattice Semiconductor as a potential acquisition target, and built the strategic and financial case for a multi-billion-pound transaction.

I led the modelling: cleaning historical financials, setting operating assumptions, building a five-year revenue, margin and cash-flow forecast, and linking CAPM-based WACC, DCF and comparable-company analysis. The model and pitchbook contributed to a Top 100 finish among more than 100 teams in the UK and Europe competition.

A companion wearable concept for detecting patterns associated with anxiety in university students. I handled market sizing, competitor mapping, product gaps and the route to the intended users.

The project forced a distinction between what a sensor might measure, what a model might infer and what a product should be allowed to claim. That uncertainty became part of the proposal rather than something to hide.