Artifact-first researcher: Indic machine-translation corpora and H100 LLM-serving benchmarks; early-stage profile
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-28) 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 publication-volume metric is available yet, so this axis is unscored. It would measure total output over time — indexed publication counts and recent productivity — which cannot be estimated from the two reviewed papers alone.
Impact—
No citation-based metric is available, so this axis is unscored. It would measure scholarly reach — citation counts, aggregate indices, and uptake of the COILD corpus or the PrefixBench-H100 artifact by later work.
Focus100/100
The bibliometric field distribution assigns all indexed records to a single field — computer systems, LLM inference serving — yielding a maximal focus score of 100. The reviewed catalog is slightly broader, adding an NLP machine-translation corpus, but both papers share a benchmark-and-resource orientation, so the concentration reading is directionally fair.
Collaboration66/100
A collaboration score of 66 rests on 26 distinct coauthors in the indexed data, with 12 recurring collaborators — including Omkar Shewale, Divakar Kumar Yadav, Nitin Kumar Mishra, and Palash Pratim Dutta — pointing to steady team-based work rather than solitary output. In both reviewed papers Kumar holds a middle, non-corresponding author position within these larger groups.
Activity—
Insufficient data exists on research cadence, so this axis is unscored. It would measure ongoing activity — publication frequency, recent-year output, and continuity — beyond the two 2026 records currently in the catalog.
Reproducibility100/100
A perfect score of 100 is the mean readiness across exactly one assessed paper (reproPapers = 1): PrefixBench-H100, rated high by Evolve as a credible single-H100 artifact. COILD is rated medium and, on current evidence, sits outside the score, so this perfect value reflects a one-paper base and reads as promising rather than established.
Research summary
Deepak Kumar is an early-stage researcher whose Evolve-reviewed work centers on building evaluation artifacts in two distinct areas: machine translation for Indic languages and systems benchmarking for LLM inference. The 2026 COILD corpus claims 1.16 million human-verified Indic-pivot parallel pairs spanning 20 language pairs, a substantial resource for low-resource Indian languages, though Evolve notes its pivot-superiority claim lacks an English-pivot control. In parallel, PrefixBench-H100 characterizes prefix reuse and time-to-first-token on a single NVIDIA H100 across vLLM and TensorRT-LLM, finding prefix-cache parity between the stacks with scheduler-driven TTFT flips; Evolve judged the artifact credible while noting thin statistics. In both papers Kumar appears as a middle, non-corresponding author within sizeable teams — 26 distinct coauthors, 12 of whom recur. Radar signals show maximal focus (100) and reproducibility (100, from one assessed paper), and a moderate collaboration score (66); output, impact, and activity axes are not yet computable. This is a name-resolved, early-stage profile built solely from OpenAlex bibliometrics and Evolve's reviewed catalog.
Collaboration pattern
Kumar works in sizeable, team-based groups: 26 distinct coauthors appear in the indexed data, with 12 recurring collaborators — Omkar Shewale, Divakar Kumar Yadav, Kshetrimayum Boynao Singh, Nitin Kumar Mishra, Palash Pratim Dutta, Atai Waris Khan, Aparna Kaushik, Avinash Kumar, Deeksha, Saroj Kumar Jha, Saloka Sengupta, and Anansa Roy. In both 2026 papers he holds a middle, non-corresponding position, consistent with contributing inside larger project teams rather than leading solo efforts.
Research keywords
Highlights
1.16M human-verified Indic-pivot MT pairs across 20 language pairs
Prefix-cache parity across vLLM and TensorRT-LLM, with scheduler-driven TTFT flips
An artifact-first, two-community output pattern
Both reviewed 2026 papers are evaluation artifacts rather than pure method papers — a translation corpus and a serving benchmark — with reproducibility readiness rated medium and high, and Evolve verdicts that pair clear value with specific methodological gaps.
Timeline
- 2026Releases COILD, an Indic-centric parallel corpus and machine-translation benchmark spanning 20 language pairs (arXiv; middle author).
- 2026Publishes PrefixBench-H100, characterizing prefix reuse and time-to-first-token in H100 LLM serving (arXiv; middle author).
- 2026Both papers enter the Evolve catalog, establishing a dual-track early profile in Indic NLP resources and LLM-serving benchmarking.
Papers in the Evolve catalog
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.
- Deepak Kumar is a common name; this profile is name-resolved from public bibliographic data, with no ORCID, institution, or country provided.
- Bibliometric signals come from OpenAlex only, and the field distribution covers a single indexed record, so radar scores rest on thin coverage.
- Output, impact, and activity dimensions are unscored because no publication counts or citation metrics were available in the input data.
- The reproducibility score of 100 is a mean over one assessed paper (reproPapers = 1) and should not be generalized.