Sachin Kumar

Sachin Kumar is a Senior Data Scientist III at LexisNexis and an independent machine-learning researcher developing practical defenses against compromised language models and more efficient approaches to AI tool use. Previously a Staff ML Engineer at Chegg, he has worked on semantic search, question answering, retrieval-augmented generation, and agentic systems. He studied at North Carolina State University from 2017 to 2019.

His independently authored research addresses two problems that become increasingly consequential as organizations customize and deploy language models:

• Sleeper-agent backdoor detection: Poisoned training data, outsourced fine-tuning, downloaded checkpoints, and insider access can implant malicious behavior that activates only under specific conditions, escaping conventional evaluations. Kumar’s backdoor-detection research proposes differential sparse autoencoders, which isolate suspicious changes by comparing internal activations before and after fine-tuning. In controlled SmolLM2-360M experiments, a year-based trigger switched SQL generation between vulnerable and parameterized implementations. His detector achieved an isolation score of 0.40, compared with approximately 0.01 for a crosscoder baseline, with perfect precision and no false positives in that setup. His AI Engineer World’s Fair presentation proposes checking model checkpoints before deployment while acknowledging significant limitations: one feature detected roughly one-quarter of triggers, the original base model is required, and larger models and adaptive attackers remain untested.

• Meta-Tool and few-shot adaptation: Kumar’s ACL research and accompanying Meta-Tool project investigate how smaller models learn unfamiliar tools. Across four benchmarks, carefully selected examples improved performance by 21.5%, structured documentation added another 5%, and a 227.8-million-parameter hypernetwork provided no measurable improvement. In those experiments, a three-billion-parameter model reached 79.7% of GPT-5’s average performance with roughly tenfold lower latency. The practical lesson: improve examples, documentation, and output constraints before introducing elaborate adaptation architectures.