Research Papers
Pratilekha: Lightweight Indic Speech Recognition
Anuran Roy, Alchemyst Labs
A family of lightweight Indic STT models - supports 12 Indian languages with sub-200ms latency on edge devices, achieving ~20% relative improvement over base Whisper on Indic benchmarks.
Is large context window all you need?
Anuran Roy, Saptarshi Pani, Arnab Sengupta, Alchemyst Labs
This paper examines the trade-offs between context sizes and latencies, highlighting the need for improved context retrieval strategies that do not bloat query sizes to the concerned Large Language Models.
BrevityBench: Measuring Intelligence per Dollar for Language Models and Agent Harnesses
Anuran Roy, Akash Kundu, Vedanta Banerjee, Khushi Mhasange, Saptarshi Pani, Debayan Pradhan
A benchmark measuring intelligence per dollar across knowledge work, coding, memory retention, and content writing — for both raw models and model-harness combinations — using atomic task checkpoints, hidden evidence-anchored rubrics, and a token-level cost ledger that attributes every dollar spent.
Listen First: Language Identity Detection with Frozen Whisper and a Learned Projector
Anuran Roy
A lightweight language identity detection system repurposing frozen Whisper encoder representations and a learned speech projector for realtime VAD-to-STT routing. 98.81% accuracy, 98.76% macro F1, 30-40ms latency on NVIDIA L40S.