verma7.com

Research

• Clustering AI Chat Transcripts at Session Level — July 2026. [pdf] [code]
  An empirical study of the topic-discovery pipeline for AI chat transcripts — session encoding, embeddings, dimensionality reduction, clustering, and LLM-based evaluation — over ~350 runs across seven datasets, including LMSYS-Chat-1M, WildChat, and Chatbot Arena. Findings: embedding quality dominates; internal metrics track ground truth only weakly; and the dominant real-world failure is “degenerate collapse,” which benchmark metrics hide. Ends with a guarded pipeline recommendation. Interactive explorer over every run and labeled cluster map.
• Co-Designing MDS Erasure Codes Across GF(2³)–GF(2⁸) — July 2026. [code]
  Minimizing the scheduled XOR count of bit-matrix erasure-code encoders. A C++ pipeline — irreducible-polynomial scan, Cauchy point optimization, exact common-subexpression scheduling, and genetic-algorithm refinement — run across six binary fields and 117 machine-verified code shapes. Finding: the smallest field that fits the code always yields the cheapest encoder. Interactive explorer over the full grid.
• Scoring Functions for Cluster Job Scheduling: 21 Policies on the Google Trace — July 2026.
  Multi-dimensional bin packing for warehouse-scale schedulers. Twenty-one placement heuristics — including the AlphaEvolve Borg heuristic and the champion of an evolutionary search — compared on the Google 2011 cluster trace across four regimes: a saturated snapshot, a live replay, Borg-style priority preemption, and the production band. Mathematical definitions, packing metrics, and cross-experiment insights. [code]
• Routing Commerce Search Queries: Simple Lookup vs. LLM Overview — July 2026. [live demo]
  A binary classifier deciding when an e-commerce search should trigger an LLM shopping overview instead of just the product grid. Five public query datasets, LLM-ensemble labeling, seven classifiers, leave-one-dataset-out generalization, and a latency/cost frontier.