Sai Pavan Kumar

—

Unclaimed · AI-structured profileOpenAlex

ML systems researcher measuring LLM inference: Pareto atlas, KV-cache tiering, disaggregated 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—

Not yet scored: the underlying productivity data — full publication counts, venues, and yearly output beyond the three 2026 arXiv preprints in this catalog — is insufficient, so no output estimate can be made.

Impact—

Not yet scored: citation counts, h-index, and downstream uptake are absent from this snapshot, so research impact cannot be measured; this axis would track citations to and influence of the reviewed papers.

Focus0/100

The focus score is 0, reflecting a field distribution split evenly at 33% each across three labeled subfields: machine learning systems (LLM inference/serving), ML systems for efficient LLM inference (memory hierarchy and caching), and ML systems (LLM inference serving). In practice these labels describe one coherent agenda — every cataloged paper targets LLM serving — so the low score appears to reflect even labeling spread across near-duplicate subfields rather than genuine topical scatter.

Collaboration36/100

The collaboration score is 36, from 5 distinct coauthors across the three cataloged papers. Recurring collaborators include Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Kapil Kumar, and Ramesh Nampelly. On all three papers the researcher appears in a middle author position rather than first or corresponding author, suggesting contribution within a small, stable team rather than an independent-led effort.

Activity—

Not yet scored: the snapshot lacks full yearly publication counts and recent-activity signals beyond the three 2026 preprints, so research tempo cannot be measured; this axis would track output cadence over time.

Reproducibility55/100

The reproducibility score is 55, aggregated from 3 reviewed papers that Evolve each rates medium on reproducibility readiness. The work is consistently measurement-based — 54 model anchors on 3 GPUs, tiered-storage simulation, and an 8-GPU routing testbed — which supports replication, but Evolve's verdicts flag unvalidated simulation, batch-one-only simulation, and missing significance tests as the main caveats.

Research summary

Sai Pavan Kumar works on machine learning systems for large language model inference. All three papers Evolve has reviewed appeared on arXiv in 2026 and sit in tightly connected sub-areas: LLM inference serving, KV-cache placement across the memory hierarchy, and disaggregated serving. The cataloged work is consistently measurement-driven. One study builds an Inference Engineering Pareto Atlas from 54 Qwen2.5-7B anchors on 3 GPUs, comparing quantization, KV-cache compression, and speculative decoding on a cost-quality-latency frontier; Evolve calls it candid and useful but notes the frontier rests on unvalidated simulation. A second simulates KV-cache placement across GPU, CPU, and SSD tiers and finds that the reported 73× long-session gain comes from added capacity rather than placement policy, with predicted-reuse and prefetch recommendations failing. A third shows a calibrated learned router achieving the best goodput for disaggregated serving on 8 A40 GPUs (0.864 versus 0.835–0.847), though at small scale across 3 traces without significance tests. All three papers carry a medium reproducibility readiness and an aggregate reproducibility score of 55. This profile is name-resolved from public bibliographic data; affiliation and citation-impact metrics are not yet available.

Collaboration pattern

Collaboration is small-team and stable: 5 distinct coauthors across the three cataloged 2026 papers, giving a collaboration score of 36. The recurring group includes Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Kapil Kumar, and Ramesh Nampelly. On every paper the researcher sits in a middle author position without a corresponding-author role, consistent with contributing measurements and analysis inside a shared ML-systems group rather than leading authorship.

Research keywords

LLM inferenceLLM servingPareto frontiercost-quality-latency tradeoffquantization (AWQ, FP8)speculative decodingKV cachememory tieringGPU HBM offloadingplacement policyprefetchingrequest routing

Highlights

Cost-quality-latency atlas of LLM inference optimizations
Builds the atlas from 54 Qwen2.5-7B anchors on 3 GPUs, comparing quantization (AWQ, FP8), KV-cache compression, and speculative decoding; Evolve finds it candid and useful but notes the frontier rests on unvalidated simulation.
KV-cache tiering gains traced to capacity, not policy
A simulation across GPU, CPU, and SSD tiers for long-lived sessions shows the 73× session gain is pure added capacity; predicted-reuse and prefetch recommendations fail. Honest, but limited to batch-one simulation.
Calibrated learned routing tops goodput for disaggregated serving
On 8 A40 GPUs the calibrated learned router reaches 0.864 goodput versus 0.835–0.847 for alternatives; a credible hardware study, though small in scale with 3 traces and no significance tests.

Timeline

2026Published the Inference Engineering Pareto Atlas, benchmarking 54 Qwen2.5-7B anchors on 3 GPUs across the cost, quality, and latency frontier.
2026Published a simulation study of KV-cache placement across GPU, CPU, and SSD, showing tiering gains for long-lived sessions come from capacity rather than placement policy.
2026Published a measured study of learned request routing for disaggregated LLM serving, with a calibrated router reaching 0.864 goodput on 8 A40 GPUs.

Papers in the Evolve catalog

Frequent co-authors

Srikanta Datta Tumkur×3 Jay Iyer×3 Mehar Simhadri×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.

Name is common; profile is name-resolved from public bibliographic data and may not represent a unique identity.
Bibliometrics come from the Evolve catalog only: 3 arXiv 2026 preprints; no institution, country, or ORCID on record.
Output, impact, and activity axes are unscored because citation and full productivity data are missing.
All Evolve verdicts describe preprints with simulation or small-scale testbeds; findings are scoped accordingly.