Hands-on work
Projects
Applied research projects that demonstrate the principles from the research in working systems, be it from physics simulation to live financial markets to real-world social networks.
Bilevel Gene Regulatory Network Inference via Surrogate-Assisted Evolutionary Ensemble Fusion
→ AUPRC 0.673 / AUROC 0.924 on DREAM4 InSilico Size100, beating published baselines PORTIA and iLSGRN
Formulated gene regulatory network inference as a bilevel optimization problem, using a surrogate-assisted evolutionary search to tune ensemble regression hyperparameters and fuse them with a causal intervention signal under strictly label-free selection criteria. Outperformed PORTIA (0.613/0.932) and iLSGRN (0.462/0.859) on precision-recall performance across independently evaluated networks.
Key components
- Bilevel formulation of GRN inference
- Surrogate-assisted evolutionary hyperparameter search
- Causal intervention signal fusion
- Label-free selection criteria
- AUPRC 0.673 / AUROC 0.924 on DREAM4 Network 1
- Outperforms PORTIA and iLSGRN baselines
Memory-Constrained Neural Architecture Search
→ 4× compression of BERT-tiny at 80.73% SST-2 accuracy, within 0.35 pp of full precision
Extended SACE-ES into a surrogate-assisted bilevel search over layer widths, bitwidths, and sparsity under a hard memory budget. Mapped a 4-point Pareto frontier across 4,096 candidate configurations using only 30 training evaluations, versus the 4,096 an exhaustive search would require, showing that surrogate-assisted optimization makes architecture search under tight memory budgets tractable.
Key components
- Surrogate-assisted bilevel search (SACE-ES)
- Joint optimization of layer widths, bitwidths, and sparsity
- Hard memory-budget constraint
- 4× compression of BERT-tiny at 80.73% SST-2 accuracy
- 30 evaluations vs. 4,096 for exhaustive search
LouvainCompress: An In-Training Neural Network Compression Mechanism
→ 93% reduction in parameters while maintaining 98.55% peak test accuracy, compressed in under 14 minutes
Developed LouvainCompress, a novel single-pass framework that uses community-detection algorithms to automatically identify and merge redundant neurons during model training. Achieved a 93% parameter reduction while maintaining 98.55% peak test accuracy on a 3.96M-parameter MLP trained on MNIST, across ten in-training compression events, making the approach highly attractive for resource-constrained enterprise deployments.
Key components
- Community-detection-based neuron merging
- Single-pass, in-training compression
- Scaled on dual RTX 5080 GPUs
- 93% parameter reduction, 98.55% peak accuracy retained
- Ten in-training compression events on a 3.96M-parameter MLP
Bilevel Memetic Swarm Intelligence for Large-Scale Max-Cut
→ Matched the literature-best known Max-Cut value of 11,624 on Stanford G-set graph G1 in 103 seconds
Designed a topology-aware memetic swarm algorithm for the Max-Cut combinatorial optimization problem, structured as a bilevel framework combining frustration-guided 1-opt local search with a degree-weighted topological gravity field for swarm consensus and a thermodynamic weakness-guided mutation operator to escape stagnation without destroying elite structure. Matched the literature-best known Max-Cut value on the 800-node, 19,176-edge G1 benchmark graph in 103 seconds.
Key components
- Bilevel memetic swarm framework
- Frustration-guided 1-opt local search
- Degree-weighted topological gravity field for swarm consensus
- Thermodynamic weakness-guided mutation operator
- Matched literature-best Max-Cut on Stanford G-set G1
Aerodynamic Shape Optimization via Genetic Algorithms
→ Minimum-drag geometries evolved from scratch using physics simulation
Applied evolutionary algorithms to optimize aircraft geometry for minimum aerodynamic drag, using Blender as the physics simulator and drag coefficient as the fitness signal. This project serves as a direct applied demonstration of the EA-surrogate paradigm central to the dissertation i.e. real-world physics replaces expensive analytical gradient computation.
Key components
- Genetic Algorithm with real-valued encoding
- Blender physics engine as fitness evaluator
- Surrogate model to reduce simulation calls
- Convergence analysis across multiple generations
Reinforced Trading Bot: C51 Distributional DQN
→ 15 of 18 backtested trades profitable over 3 months, averaging 8% profit per trade
Trained a Categorical Deep Q-Network (C51) on live candlestick market data. The agent autonomously learned when to trigger stop-losses, execute buy/sell signals, and switch between long and short positions, with zero human-labelled training data. Backtested over 3 years of data, executing 15 of 18 total trades over a 3-month period profitably, demonstrating that distributional RL handles the high reward variance of financial markets better than standard DQN.
Key components
- C51 distributional return modelling
- Live OHLCV candlestick data pipeline
- Zero human-labelled training data
- Custom Gymnasium trading environment
- 15 of 18 backtested trades profitable, ~8% average profit
Information Cascade Prediction via Graph Embeddings
→ DFS-biased node2vec outperforms state-of-the-art CasFlow on cascade prediction
Investigated whether depth-first or breadth-first structural priors produce better embeddings for predicting information cascade popularity. DFS-biased node2vec consistently outperforms the state-of-the-art CasFlow model thus establishing that deep structural exploration captures diffusion dynamics better than breadth-first neighbourhood aggregation on social network cascade data.
Key components
- DFS-biased node2vec embedding
- Comparison vs CasFlow baseline
- Cascade popularity as regression target
- Real-world social network datasets
- Depth-first prior as key inductive bias