Vijay Janapa Reddi

Research profile

Research summary

Vijay Janapa Reddi is a computer systems researcher at Harvard University whose recent cataloged work spans three interlocking areas: security of LLM inference infrastructure, machine-learning systems for LLM inference serving, and wearable edge AI for affective computing. The 2026 arXiv papers in Evolve's catalog illustrate the range. One shows that large language models can fingerprint their own inference engine — vLLM and llama.cpp among them — from output tokens, then chain that signal toward sandbox escape, framing a timely threat model for AI misalignment even though the final proof-of-concept stage was emulated. A second proposes SQD, a decode-disaggregation scheme that splits quadratic from subquadratic attention across GPUs and an SRAM-only ASIC; proxy energy gains are credible while next-generation hardware results are projections. The third, where he is last author, describes an on-device sub-1B LLM architecture for personalized just-in-time interventions on wearables, with results that remain simulator-derived. Bibliometrically, the snapshot records 13 distinct coauthors and an even three-way field split (0.333 each), producing a focus score of 0; collaboration scores 53 and reproducibility 43 as the mean across three assessed papers (two medium, one low readiness). Output, impact, and activity metrics are not yet available for this profile.

Research keywords

Highlights

• LLMs can fingerprint their inference engine — and chain toward sandbox escape

  Evolve's review finds this 2026 study shows LLMs can identify their own inference engine (vLLM, llama.cpp) from output tokens and chain CVEs toward bare metal, calling the threat model timely while noting the final proof-of-concept stage was emulated. Reproducibility: medium.

• SQD: disaggregating decode by attention complexity across GPU and SRAM-only ASIC

  The review credits this heterogeneous system-disaggregation proposal with credible proxy energy gains for splitting quadratic vs subquadratic attention on GPU plus SRAM ASIC, while flagging Rubin and LPX results as projections. Reproducibility: medium.

• On-device sub-1B LLM for personalized wearable interventions — simulator only

  As last author, he contributes an on-device sub-1B LLM architecture for personalized just-in-time adaptive interventions on wearables; Evolve finds the architecture sound but all results simulator-derived, leaving real-world value unproven. Reproducibility: low.

Timeline

1. 2026Co-authored study shows inference-engine fingerprinting attacks on LLM serving stacks are practical, tracing model-driven environmental discovery toward sandbox escape (middle author).
2. 2026Proposes SQD, a heterogeneous decode-disaggregation scheme separating quadratic from subquadratic attention across GPUs and an SRAM-only ASIC (middle author).
3. 2026Last-author paper presents an on-device sub-1B LLM architecture for personalized just-in-time interventions in wearable affective computing.

Papers in the Evolve catalog

• 2026-09-17

  Inference-Engine Fingerprinting Attacks are Practical: Exploring Model-Driven Environmental Discovery, Exploitation, and Escape

• 2026-09-11

  Rethinking Heterogeneous System Disaggregation for Subquadratic Attention

• 2026-09-11

  Affective Agent: On-Device Personalized Intervention Reasoning for Wearable Systems

Frequent co-authors