02 / Industry collaboration · Applied AI research
Unisys Research.
A research collaboration exploring reward modeling, pairwise ranking and semantic retrieval for specialized programming-language datasets.
01 / OVERVIEW
The context.
This industry-collaborative research project focuses on data-efficient reward modeling for proprietary ALGOL datasets. The work combines structured dataset preparation, pairwise comparisons, document-processing pipelines and retrieval-based evaluation.
02 / CONTRIBUTIONS
What I worked on.
- Developed a data-efficient reward-modeling pipeline using pairwise ranking techniques.
- Curated structured datasets containing 600+ samples and 3,000+ pairwise comparisons through filtering and augmentation.
- Engineered document-processing and FAISS-based semantic retrieval pipelines involving cleaning, chunking, filtering and indexing.
- Evaluated model performance using pairwise and groupwise accuracy metrics with confidence-estimation techniques.
03 / ENGINEERING NOTES
How it comes together.
01The work connects data curation, preference comparisons and retrieval infrastructure as components of a broader model-evaluation pipeline.
02The portfolio intentionally describes the work at a high level. Proprietary datasets, internal implementation details and confidential research material are not included.
04 / TECHNOLOGY