Accessibility-aware urban routing with .NET 9, PostGIS, Redis/Kafka workers, graph preprocessing, hazard scoring, and an Expo mobile client.

Reproducible partner-aware antibody-antigen interface modeling research

3.238B dense pretraining through 100B tokens: measured checkpoints, implementation, training evidence and reproducible result pages.

Advanced research engineering portfolio for Yin Li: LLM systems, evaluation, post-training, retrieval, and agent traces.

Executable-code evaluation and verifier-guided inference, with reproducible benchmarks and separate model-weight retention audits.

From-scratch 135M Transformer pretraining on 10B FineWeb-Edu tokens using a single NVIDIA L20, with public checkpoint and lm-eval comparisons.

A 1.1B language model trained from scratch on one L20, with frozen evaluation records and a separate multimodal research track.

Benchmark toolkit for quantized LLM quality retention, throughput, latency, and sustained availability.

Reproducible NL2SQL fine-tuning and multi-path inference benchmark on Spider/BIRD using Qwen2.5-Coder-7B and one NVIDIA L20 GPU.

Evidence-gated multimodal protein pretraining and evaluation

Stateful agent-interface benchmark for rewrite vs stable-line patch reliability in ordering agents.

Evidence-first collection of from-scratch language models trained on a single NVIDIA L20

Active web-search RAG workbench with provider routing, source extraction, citation verification, and extractive fallback.

LLM inference systems research: request geometry, runtime profiling, and serving changes with reproducible GPU evidence.