Melbourne, Australia
Will Lienert
Physics graduate who builds machine-learning systems end to end — from the data pipeline underneath to the model doing the work.
Try··
SCROLL TO SEE MY PROJECTS ↓
News that assembles itself around you.
A news platform built end to end. It scrapes and geocodes sources, then groups documents into events by extracting named entities and matching them with a source-aware TF-IDF, so an outlet's boilerplate carries no weight. Events are scored on source credibility, LLM-judged importance, interest and locality, then ranked against a vector of what you care about and where you live.
- sources
- ingest
- cluster
- rank
- publish
Live from production · anonymous feed
Predicting chaos with ten qubits.
Reservoir computing trains nothing but a linear readout — so the reservoir can be any complex dynamical system, including a quantum one. I simulated a ten-qubit transverse-field Ising model in QuTiP and used it to forecast a chaotic time series, with a split-permutation classical memory separating single-site from co-occurrence observables.
A language model, from the tensor up.
A GPT-2-scale decoder written from scratch in PyTorch — no model libraries, just the architecture and a training loop — pre-trained on FineWeb-Edu and then instruction-tuned.
- Architecture
- 12 blocks, 12 heads, 768-dim residual stream, 1024-token context, tied embeddings, flash attention. ~124M parameters.
- Training loop
- Hand-rolled: 20k steps at ~500k tokens per batch, cosine 6e-4 → 6e-5 after 715 warmup steps, checkpoint and resume.
- Data
- A 2.5B-token FineWeb-Edu shard tokenised to a memory-mapped binary, then No Robots for instruction tuning.
- Readout
- Micro-batch size auto-tuned to the GPU, gradient accumulation to hit the token budget, all of it containerised.
Status
Deploying trained model soon...
Available now
Looking for a graduate or junior role where I can keep building things like these.
BSc Physics, RMIT (2025). Melbourne-based, open to hybrid or in-office. This site is a Next.js app; the chat is retrieval-augmented over a knowledge base I maintain by hand.