Mehar Simhadri

—

Unclaimed · AI-structured profileOpenAlex

ML systems researcher measuring LLM inference: Pareto frontiers, KV cache tiering, and learned request routing

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

The output axis is not scored: no total publication count, venue list, or yearly output volume is available in the current snapshot. It would measure production volume and consistency across venues once a fuller publication record exists.

Impact—

Impact is unscored because no citation counts, h-index, or usage signals exist in the input data. This axis would measure how far the three catalogued inference-systems papers propagate through citations and downstream adoption.

Focus0/100

Focus scores 0 because the field distribution splits evenly across three labels — machine learning systems (LLM inference/serving), ML systems for efficient inference (memory hierarchy and caching), and ML systems (LLM inference serving) — each at one third. The labels overlap heavily in substance, so the low score reflects labeling spread rather than genuine topic dispersion; the work coheres around LLM inference measurement and systems evaluation.

Collaboration36/100

Collaboration scores 36 on 5 distinct coauthors across three papers, a small and stable pool. Recurring names include Srikanta Datta Tumkur, Jay Iyer, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly, suggesting a compact working group. All catalogued papers place this researcher in a middle author position, so the record shows steady teamwork rather than visible independent leadership.

Activity—

The activity axis is unscored: yearly publication volume and cadence cannot be computed from a snapshot containing only three 2026 papers. It would measure publishing rhythm and sustained engagement once multi-year records are available.

Reproducibility55/100

Reproducibility scores 55, the mean across 3 reviewed papers, each rated medium readiness. The pattern is consistent: all three studies state their setups plainly — 54 anchors on 3 GPUs, batch-one simulation, 8 A40s with 3 traces — but each also carries stated limits such as unvalidated simulation and no significance testing, which keeps readiness at medium rather than high.

Research summary

Mehar Simhadri is an early-stage researcher working in machine learning systems, with a catalogued record concentrated entirely on LLM inference and serving. All three papers Evolve has reviewed appeared in 2026 on arXiv, and each takes a measurement-first stance: an Inference Engineering Pareto Atlas benchmarking 54 Qwen2.5-7B optimization anchors across 3 GPUs to map cost-quality-latency tradeoffs; a simulation study of KV cache placement across GPU, CPU, and SSD tiers for long-lived sessions; and a hardware study of learned request routing for disaggregated serving, where a calibrated router reached 0.864 goodput on 8 A40 GPUs against 0.835-0.847 for baselines. The field distribution is an even three-way split across machine learning systems labels that all describe LLM inference, which yields a focus score of 0 despite the topical coherence. Collaboration metrics are modest: 5 distinct coauthors and a collaboration score of 36, with repeated partnerships including Srikanta Datta Tumkur, Jay Iyer, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly. Reproducibility readiness is rated medium for all three papers, producing a reproducibility mean of 55. Output, impact, and activity axes are not yet scored because the underlying bibliometric data is insufficient.

Collaboration pattern

Across three 2026 arXiv papers, Mehar Simhadri works within a compact, consistent team: 5 distinct coauthors in total, yielding a collaboration score of 36. The recurring group spans Srikanta Datta Tumkur, Jay Iyer, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly, indicating repeated joint work rather than one-off partnerships. All catalogued papers list a middle author position without corresponding status, so the visible pattern is steady contribution inside a small group rather than solo leadership.

Research keywords

LLM inferenceLLM servingKV cachememory tieringGPU HBM offloadingplacement policyprefetchingprefill-decode disaggregationrequest routingquantization (AWQ, FP8)speculative decodingcost-quality-latency tradeoff

Highlights

Mapping the inference optimization frontier
The Inference Engineering Pareto Atlas measures 54 Qwen2.5-7B anchors on 3 GPUs to chart cost-quality-latency tradeoffs; Evolve calls it candid and useful but notes the frontier rests on unvalidated simulation.
Testing whether KV cache tiering policy matters
A simulation of KV cache placement across GPU, CPU, and SSD finds the 73× session gain is pure capacity, not policy, and that predicted-reuse and prefetch recommendations fail — honest, but limited to batch-one simulation.
Calibrated learned routing for disaggregated serving
On 8 A40 GPUs, a calibrated learned router tops goodput at 0.864 versus 0.835-0.847 for baselines; Evolve rates it credible but flags small scale, 3 traces, and no significance tests.

Timeline

2026Published 'Calibrate, Then Route' on arXiv, a measured study of learned request routing for disaggregated LLM serving, with a calibrated router reaching 0.864 goodput on 8 A40 GPUs.
2026Published 'Where Should the KV Cache Live?' on arXiv, a simulation of KV cache placement policies across GPU, CPU, and SSD for long-lived sessions.
2026Published 'The Inference Engineering Pareto Atlas' on arXiv, measuring 54 Qwen2.5-7B optimization anchors on 3 GPUs across cost, quality, and latency.

Papers in the Evolve catalog

Frequent co-authors

Srikanta Datta Tumkur×3 Jay Iyer×3 Sai Pavan Kumar×3 Sai Kapil Kumar×3 Ramesh Nampelly×3

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Radar scores are computed from a thin bibliometric snapshot; output, impact, and activity axes lack sufficient underlying data and are left unscored.
No institution, country, or ORCID is available; the profile is name-resolved from public bibliographic data, and name ambiguity is possible.
Evidence is limited to three 2026 arXiv papers reviewed by Evolve; citation counts, h-index, and peer-reviewed venue history are not in scope.
All catalogued papers list a middle author position with no corresponding-author role; authorship-order conventions across venues are unknown.