Competitive ML

Problem: Classify 28×28 RGB images using only NumPy and pandas, without standard machine-learning libraries, under a two-hour limit and a maximum of two submissions.

Problem: Build an AI system that can understand and solve mathematics problems written in Bengali.

Solution: Math101 combined Block Decomposition with dominance- and threshold-based pruning to reduce redundant inference. The reported system achieved 4× faster inference while maintaining 84% overall accuracy.

Problem: Detect dataset references in research papers and determine how each referenced dataset is used.

Solution: Used regex candidate extraction, DeBERTa filtering, and Qwen-based classification, with a separate DOI pipeline for noisy references and role classification.

Problem: Answer questions about STEM diagrams and problem illustrations using vision-language models under competition constraints.

Problem: Generate compact SVG illustrations from natural-language descriptions under strict runtime and file-size constraints.

Solution: Used SDXL Flash for image generation, custom and VTracer-based image-to-SVG pipelines, a fine-tuned SmolLM2 for TIFA-style questions, and PaliGemma for candidate selection.

Problem: Solve national-level mathematical olympiad problems with open-weight AI models under competition compute constraints.

Solution: Built an LLM inference pipeline that solved 28 of 50 public problems and 26 of 50 private problems.