02 / Projects
Research artifacts, not products.
Each of these exists to answer a question rather than to be used by anyone. They are written to be re-run: every configuration reports accuracy, latency, and memory together, measured on the same machine under the same protocol.
RTX 4060 Laptop · CUDA · 8GB VRAM
Featured project
QuantLab
Hardware-aware mixed-precision quantization framework for transformer architectures.
Research question
Given a fixed hardware budget, which parts of a transformer should keep their bits?
QuantLab is a research-oriented framework for studying post-training quantization of transformers when precision is allowed to vary across the model. It measures how sensitive each component is, proposes allocations under a budget, benchmarks them on a real GPU, and keeps the configurations that nothing else dominates.
The design principle is that no allocation counts as good until its cost has been measured. A format that has no efficient kernel is smaller on paper and slower in practice, and the framework is built to make that visible rather than to hide it behind a compression ratio.
Fig. 01 — Configuration space
Method
- 01
Layer-wise sensitivity analysis
Quantize one component at a time and record what quality it costs, building a profile across depth and across matrices.
- 02
Post-training quantization
Calibrate ranges from a small representative sample; no retraining, so the method stays usable on a single consumer GPU.
- 03
Mixed-precision allocation
Assign INT4, NF4, INT8, or FP16 per component under a total budget, guided by the measured profile.
- 04
Hardware benchmarking
Time and measure every candidate on the target GPU with warm-up, synchronisation, and tail latency recorded.
- 05
Multi-objective search
NSGA-II over allocations, returning a Pareto front across accuracy, latency, and memory instead of a single winner.
Fig. 02 — Hardware landscape
FP16 baseline
Mixed precision
| Allocation | FP16 | Mixed |
|---|---|---|
| Weights | 4.40 | 1.55 |
| KV cache | 1.05 | 0.60 |
| Activations | 0.65 | 0.65 |
| Runtime context | 0.60 | 0.60 |
| Headroom | 1.30 | 4.60 |
Gigabytes · 8 GB capacity
- Status
- Active · research project
- Year
- 2025 — present
- Hardware
- NVIDIA RTX 4060 Laptop GPU — 8GB VRAM
- Stack
- Python · PyTorch · CUDA · Hugging Face Transformers · NumPy · Linux
- Concepts
- Post-training quantization · layer sensitivity · mixed precision · NF4 · precision allocation · NSGA-II · Pareto optimization
- Reporting
- Accuracy, P50/P95 latency, peak memory, and model size per configuration
Results are not summarised here as headline numbers. The framework produces a front of configurations rather than a single best model, and the honest summary of a run is the shape of that front on the hardware it was measured on.
Fig. 03 — Allocation pattern
- FP16
- INT8
- INT4
Featured project
Explainable Multimodal Medical AI
Skin cancer risk stratification from images and clinical metadata, with interpretability and uncertainty treated as results.
Research question
If a model is going to be trusted in a clinical setting, what has to be reported alongside its prediction?
A dermoscopic image alone leaves out what a clinician would obviously use: the patient’s age, sex, and where on the body the lesion is. This project combines both, and then spends most of its effort on the harder half of the problem — establishing whether the resulting predictions can be trusted.
Attribution shows which pixels and which clinical variables drove a decision. Monte Carlo dropout gives a measure of predictive uncertainty. Temperature scaling makes the reported confidence mean something. Validation splits by patient rather than by image, and error rates are reported per demographic group instead of only in aggregate.
EfficientNet-B0·Multimodal fusion·Grad-CAM++·SHAP·MC dropout·Temperature scaling
Fig. 04 — Fusion pipeline
01 · Input
Dermoscopic image
Lesion photograph, resized and normalized.
Clinical metadata
Age, sex, anatomical site — structured, sparse, occasionally missing.
02 · Encode
EfficientNet-B0
Image embedding from a compact convolutional backbone.
Metadata encoder
Clinical variables projected into the same representation space.
03 · Fuse & verify
Multimodal fusion → risk stratification
A joint representation produces the risk output.
Grad-CAM++ · SHAP
Attribution over pixels and over clinical variables.
MC dropout · temperature scaling
Predictive uncertainty, then calibration of the reported confidence.
Patient-level split · demographic audit
Validation that does not leak patients, and error rates reported per group.
- Status
- Completed · research project
- Year
- 2025 — 2026
- Modalities
- Dermoscopic images · structured clinical metadata
- Stack
- Python · PyTorch · EfficientNet · scikit-learn · SHAP · FastAPI · Streamlit
- Reliability
- Uncertainty estimation · calibration · patient-level validation · demographic fairness audit
What I took from it
That reliability work and efficiency work are the same kind of work. Both are about being precise regarding what a model actually gives you, and both are undermined by reporting a single number. It is the reason calibration and fairness stayed on my list when my focus moved to quantization.
03 / Planned work
Not started · scoped only
BitFlex — Adaptive Precision Transformer Inference
PlannedNo implementation yet
Mixed precision, as I am currently studying it, is decided once and then fixed. BitFlex is the question of whether that decision has to be static at all.
Research question
Can transformer inference select numerical precision at runtime, according to layer sensitivity, input characteristics, and hardware constraints?
Dimensions to compare
- 01
Static mixed precision
One allocation, fixed ahead of deployment. The baseline.
- 02
Component-aware precision
Allocation varying by component type rather than only by depth.
- 03
Input-adaptive precision
Precision chosen per input, on the assumption that easy inputs need fewer bits.
- 04
Hardware-aware adaptive precision
Runtime selection constrained by what the device can actually execute quickly.
Intended evaluation
Accuracy·F1·P50 latency·P95 latency·Throughput·VRAM·Model size·GPU utilisation
04 / Ongoing
Maintained continuously
Smaller work kept in the open.
AI & Machine Learning Portfolio
2025 — Present
A running collection of machine learning experiments — data processing, model development, training, evaluation, and analysis — kept as a record of method rather than of results.
Developer & Research Portfolio
2026 — Present
This site: research interests, project write-ups, working notes, and technical work, maintained as its own small engineering exercise.
05 / Also planned
Further directions, listed honestly as intentions.
EdgeFormer
Later
Hardware-aware transformer optimization for constrained devices
Carrying the same allocation question down to devices where memory bandwidth, not compute, is the binding constraint.
ParetoML
Later
Accuracy–efficiency optimization tooling
Making multi-objective model selection a routine, reportable step instead of an ad-hoc comparison of two checkpoints.
AutoQuant
Later
Automated transformer compression
Automating the search over compression decisions so the method transfers to models and accelerators I do not own.