Ankan Biswas

About Ankan Biswas

I just love numbers.

Experience

1. Ai Research Scientist

   Fractal

   Aug 2023 — Present · Gurugram, IN

   Researcher, Fathom-R1-14B — Open-Source STEM Reasoning LLM:– Designed and post-trained a reasoning LLM (Fathom-R1-14B), achieving state-of-the-art performance on olympiad-level benchmarks (AIME-25, HMMT-25, etc.) within a 16K context window and a $499 post-training budget.– Surpassed frontier models like o3-mini, o1-mini, etc. in pass@1 accuracy and rivaled o4-mini-low in cons@64 scores.– Achieved a perfect score (100%) on IITJEE 2025 Mathematics and trended #5 on HuggingFace with 20K+ downloads.• Silver Medalist, AI Mathematical Olympiad (AIMO) Progress Prize 1:– Secured a silver medal (98th percentile) in the globally recognized AIMO Kaggle competition, solving olympiad-level math problems at scale using LLMs within strict compute constraints.– Engineered a hybrid Chain-of-Thought (CoT)+ Program-of-Thought (PoT) reasoning framework, generating multiple solution paths and applying majority voting to improve answer robustness.– Implemented ML-based classification techniques to automatically distinguish correct from incorrect reasoning chains, significantly boosting solution accuracy and enabling scalable AI-driven quantitative problem-solving.

Education

• Indian Institute of Technology, Kharagpur

  Integrated Dual Degree (B.Tech + M.Tech), Electrical and Electronics Engineering