ML-systems researcher on efficient LLM inference: KV-cache placement, learned routing, cost-quality-latency trade-offs
3
papers here
3
reviewed by Evolve
0
first-author
0
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—
No score is reported for output because publication counts and productivity metrics are unavailable in the current snapshot. This axis would measure the volume, venues, and trajectory of this researcher's publications over time.
Impact—
Impact is unscored because citation counts and influence indicators are not present in the current data. This axis would measure how widely this researcher's papers are cited and adopted by the wider research community.
Focus0/100
The radar reports a focus score of 0. The underlying field distribution splits the three cataloged papers evenly across three closely related subfields — machine learning systems for LLM inference/serving, ML systems for efficient inference (memory hierarchy and caching), and ML systems for LLM inference serving — each at 0.333. Read together they form a tight cluster around one theme; the 0 reflects no single dominant field label rather than scattered interests.
Collaboration36/100
Collaboration scores 36, based on 5 distinct coauthors. All three Evolve-cataloged 2026 papers list this researcher as a middle author and none as corresponding author. The recurring coauthor pool includes Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, and Ramesh Nampelly, suggesting a small stable team; the score sits below mid-range, consistent with a compact collaboration footprint.
Activity—
Activity is unscored because recent publication cadence data is insufficient in the current snapshot. This axis would measure publishing rhythm and continuity, for example the number of papers per year over recent years.
Reproducibility55/100
Reproducibility scores 55, the mean readiness across all 3 cataloged papers, each rated medium. All three describe concrete setups — 54 Qwen2.5-7B anchors on 3 GPUs, GPU/CPU/SSD tiering simulation, and 8 A40 routing experiments — but Evolve flags unvalidated simulation, batch-one scope, and absent significance tests, keeping readiness below the top band.
Research summary
Sai Kapil Kumar is a machine-learning-systems researcher whose Evolve-reviewed work concentrates on engineering efficient large language model (LLM) inference and serving. The catalog lists three 2026 arXiv papers, in each of which Kumar is a middle author and none a corresponding author. The first, The Inference Engineering Pareto Atlas, measures 54 Qwen2.5-7B anchors on 3 GPUs and simulates a cost-quality-latency atlas spanning quantization (AWQ, FP8), KV-cache compression, and speculative decoding; Evolve calls it candid and useful but notes the frontier rests on unvalidated simulation. The second asks where the KV cache should live across GPU, CPU, and SSD, finding that a 73x session gain comes purely from added capacity rather than placement policy, with predicted-reuse and prefetch recommendations failing under batch-one simulation. The third, Calibrate, Then Route, shows a calibrated learned router topping goodput on 8 A40 GPUs (0.864 versus 0.835-0.847), though at small scale across 3 traces without significance tests. Radar scores give reproducibility 55 (mean across 3 papers), collaboration 36 with 5 distinct coauthors, and a focus score of 0 from an even three-way field split; output, impact, and activity lack data and remain unscored.
Collaboration pattern
Kumar's cataloged work shows a compact collaboration pattern: 5 distinct coauthors, with a recurring core of Srikanta Datta Tumkur, Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, and Ramesh Nampelly. In all three 2026 arXiv papers Kumar appears as a middle author rather than first or corresponding author, fitting a contributor role within a small, stable ML-systems team rather than an independent group lead.
Research keywords
Highlights
Atlas of inference optimizations across cost, quality, and latency
KV-cache tiering: capacity, not policy, drives the gains
Calibrated learned routing tops goodput for disaggregated serving
Timeline
- 2026Published The Inference Engineering Pareto Atlas, measuring 54 Qwen2.5-7B anchors on 3 GPUs across quantization, KV-cache compression, and speculative decoding.
- 2026Published the KV-cache placement study across GPU, CPU, and SSD, showing that a 73x session gain comes from capacity rather than placement policy.
- 2026Published Calibrate, Then Route, a measured study of learned request routing for disaggregated LLM serving on 8 A40 GPUs.
Papers in the Evolve catalog
The Inference Engineering Pareto Atlas: Which Optimizations Dominate the Cost, Quality, and Latency Frontier?
Where Should the KV Cache Live? Placement Policies Across GPU, CPU, and SSD for Long-Lived Sessions
Calibrate, Then Route: A Measured Study of Learned Request Routing for Disaggregated LLM Serving
Frequent co-authors
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.
- Scores derive from Evolve's bibliometric snapshot; output, impact, and activity are unscored because the underlying metrics are missing.
- The profile rests on only 3 arXiv preprints (all 2026) reviewed in the Evolve catalog; no peer-reviewed venue data is included.
- Institution, country, and ORCID are not provided; Sai Kapil Kumar is a relatively common name, so this profile is name-resolved from public bibliographic data.
- Radar scores (collaboration 36, focus 0, reproducibility 55) are platform-computed indices, not direct measurements.