Research
Research areas and published results.
We deploy software, measure failures, build benchmarks, and publish the methods and limitations.
Collaborative scientific writing
How can a browser editor combine conflict-free collaboration, full TeX compilation, provenance, and an AI assistant that validates edits by compiling them?
02Code-generated scientific media
Can structured animation code produce more controllable STEM explanations than end-to-end pixel generation?
03Persistent agent context
How should agents retrieve relationships, temporal updates, and project knowledge across sessions without flattening everything into vectors?
04Computing in local language
What changes when a learner can express programming logic in Bangla while retaining a path into conventional syntax?
Benchmark spotlight
Context Heavy on LoCoMo retrieval.
A bounded two-conversation, 304-question slice, evaluated at k=10 with neutral titles. Context Heavy led the compared systems on retrieval quality; its remote reranker also made it slower. Both facts matter.
| System | Recall@10 | nDCG@10 | p50 |
|---|---|---|---|
| Context Heavy | 71.2% | 60.1% | 1,708 ms |
| supermemory | 63.3% | 45.4% | 713 ms |
| gbrain | 58.0% | 38.2% | 3,071 ms |
| mem0 | 20.6% | 17.4% | 1,184 ms |
Retrieval-only comparison; not comparable with answer-judge leaderboards. Full LoCoMo is larger. Results and reproduction notes are published in CH-Bench.
Publication
Synergistic application of advanced machine learning and computer vision techniques for the detection of exoplanet and star
A 2024 BRAC University research work co-authored by Ashraful Kabir Alif and Shihab Shahriar Antor with S. J. Hossain, M. I. Hossain, and A. P. Roy. The work studies contrastive and few-shot learning for faint-object detection in JWST imagery and reports 92.16% accuracy with YOLOv5.