Srikanta Datta Tumkur

—

Unclaimed · AI-structured profileOpenAlex

LLM inference systems researcher; first author of a cost-quality-latency Pareto atlas of inference optimizations

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—

No output score is available: the catalog holds a single paper, too little to measure publication volume or rate. This axis would count indexed outputs — preprints, conference and journal papers — over time.

Impact—

No impact score is available: with one 2026 entry there are too few citations or adoption signals to judge influence. This axis would track citations, downstream reuse, and references from other indexed work.

Focus0/100

Focus scores 100 because the field distribution is undifferentiated: the only cataloged paper, and thus 100% of the indexed output, sits in machine learning systems (LLM inference/serving). There is no second field to dilute the profile. The score should still be read cautiously — it rests on a single 2026 arXiv paper — but the chosen problem, mapping inference optimizations onto a cost-quality-latency frontier, is a sharply defined slice of ML systems rather than a spread across topics.

Collaboration36/100

Collaboration scores 36, computed from 5 distinct coauthors in the current bibliometric snapshot. The frequent-collaborator list — Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly — shows the same small cluster recurring, pointing to a stable in-house team rather than an extended network of external partners. With only one cataloged paper, this is a thin base: the score describes team size today, not an established collaboration pattern.

Activity—

No activity score is available: a single 2026 entry cannot establish a publishing rhythm. This axis would measure recency and cadence — how consistently new work appears year over year.

Reproducibility55/100

Reproducibility averages 55 over the single paper assessed so far, so the sample is exactly one study. The Pareto Atlas is rated medium reproducibility readiness: its 54-configuration measurement of Qwen2.5-7B on 3 GPUs is concrete and candidly reported, but Evolve's review cautions that the headline frontier rests on unvalidated simulation, meaning others re-running the measurements may not land on the same frontier without first validating the simulation layer.

Research summary

Srikanta Datta Tumkur is an early-stage researcher in machine learning systems, specializing in large language model inference and serving. This profile is name-resolved from public bibliographic data; no institutional affiliation, country, or ORCID is on record. The Evolve catalog holds one paper: the 2026 arXiv preprint 'The Inference Engineering Pareto Atlas: Which Optimizations Dominate the Cost, Quality, and Latency Frontier?', on which Tumkur is first author. The study measures 54 Qwen2.5-7B serving configurations on a three-GPU testbed and organizes them into a cost-quality-latency atlas covering quantization (AWQ, FP8), KV cache compression, and speculative decoding. Evolve's review calls the work candid and useful while flagging that its Pareto frontier rests on unvalidated simulation. The bibliometric radar reflects this early stage: a perfect focus score of 100, since all cataloged output sits in machine learning systems (LLM inference/serving); a modest collaboration score of 36 from five distinct coauthors; and a reproducibility mean of 55 across one assessed paper, rated medium readiness. Output, impact, and activity axes are unscored for lack of data. Recurring coauthors — Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly — suggest a small, stable team. Treat this as an early-stage profile that will firm up as more output is indexed.

Collaboration pattern

The collaboration footprint is compact: 5 distinct coauthors appear in the snapshot, and the same five names recur as frequent collaborators — Jay Iyer, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly. That pattern points to a small, stable team working on the same measurement agenda, with Tumkur in the first-author seat. The collaboration axis scores 36, consistent with a narrow but consistent co-authorship base rather than a broad network.

Research keywords

LLM inferenceLLM serving systemsPareto frontier analysiscost-quality-latency tradeoffquantization (AWQ, FP8)KV cache compressionspeculative decodingempirical benchmarkinglatency optimizationmachine learning systems

Highlights

First-author Pareto atlas of LLM inference optimizations
In the 2026 arXiv preprint 'The Inference Engineering Pareto Atlas', Tumkur leads a systematic comparison of the major inference optimizations — quantization (AWQ, FP8), KV cache compression, and speculative decoding — mapped onto a single cost-quality-latency frontier.
54 configurations measured on a three-GPU testbed
The atlas is grounded in measurements of 54 Qwen2.5-7B serving anchors run on 3 GPUs, giving the optimization comparison an empirical base instead of relying on modeled numbers alone.
Candid about the limits of simulation
Evolve's review calls the paper candid and useful while flagging that the reported frontier rests on unvalidated simulation; the profile carries this limitation as a caveat and rates the paper's reproducibility readiness medium.

Timeline

2026Posts the arXiv preprint 'The Inference Engineering Pareto Atlas' as first author
2026Paper enters the Evolve catalog under machine learning systems (LLM inference/serving)
2026Evolve review credits measurements of 54 Qwen2.5-7B anchors on 3 GPUs and rates reproducibility readiness medium

Papers in the Evolve catalog

2026-09-17first author
The Inference Engineering Pareto Atlas: Which Optimizations Dominate the Cost, Quality, and Latency Frontier?
2026-09-14first author
Where Should the KV Cache Live? Placement Policies Across GPU, CPU, and SSD for Long-Lived Sessions
2026-09-14first author
Calibrate, Then Route: A Measured Study of Learned Request Routing for Disaggregated LLM Serving

Frequent co-authors

Jay Iyer×3 Mehar Simhadri×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.

The bibliometric sample is one cataloged paper; output, impact, and activity dimensions are unscored for lack of data.
No institutional affiliation, country, or ORCID is on record; the profile is name-resolved from public bibliographic data and namesakes cannot be ruled out.
Focus, collaboration, and reproducibility scores rest on a single paper and 5 distinct coauthors, so treat them as preliminary.
Evolve's verdict flags that the paper's frontier rests on unvalidated simulation — a limitation of the underlying study, reported here without adjustment.