Direction

Research

I'm an engineer first, but the engineering aims at research questions: how transformers represent language, how to serve them efficiently, and how to ground their outputs in verifiable sources.

01Interests

Currently exploring

Transformer Architectures

Attention variants, positional encodings, normalization schemes, and the design space of encoder and decoder stacks.

Efficient Inference

KV caching, speculative decoding, quantization, and the systems work that makes large models cheap to serve.

Retrieval & Grounding

Dense retrieval, retrieval-augmented generation, and grounding generative models in verifiable sources.

Representation Learning

What embeddings encode, how pretraining objectives shape them, and how to evaluate them honestly.

Generative AI

Decoder-only language modeling, sampling strategies, and evaluation of open-ended generation.

NLP

Tokenization, pretraining pipelines, and task-specific finetuning across the modern NLP stack.

02Reading

Papers behind the code

The references my implementations are built against — each one maps to a concrete component.

Attention Is All You Need
Vaswani et al., 2017
The reference point for every encoder and decoder I've implemented.
RoFormer: Enhanced Transformer with Rotary Position Embedding
Su et al., 2021
Basis for Zenith's positional encoding, including its behaviour under KV caching.
Fast Inference from Transformers via Speculative Decoding
Leviathan et al., 2023
Basis for Zenith's greedy-exact speculative decoding path.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis et al., 2020
The lineage Meridian's grounded pipeline builds on.
Training Compute-Optimal Large Language Models
Hoffmann et al., 2022
Scaling laws, and how to budget compute honestly at laptop scale.

03Roadmap

Where the work is going

Now
Close Meridian's measured domain-transfer gap: SciNLI adaptation for the NLI verifier, whose 78.3% SNLI accuracy does not yet carry over to biomedical prose.
Next
Scale Meridian past laptop constraints — the ~200K-abstract PubMed corpus and the cited-answer generator trained at scale — plus QLoRA and FSDP in Zenith.
Later
Toward a broader grounded biomedical platform: richer sources beyond abstracts, stronger calibration, and human-agreement studies for the verifier.

04Output

Publications & talks

Publications

Nothing published yet — this grows as the roadmap matures. Engineering write-ups live in the notes in the meantime.

Talks

None scheduled. If you'd like a walkthrough of from-scratch NLP systems — tokenizers to serving — get in touch.