Suvradip Paul
National Payments Corporation of India · India
NLP researcher at NPCI building and benchmarking domain-specific LLM agents for Indian retail banking
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
Insufficient data: no reliable publication-count or venue record is available beyond the OpenAlex snapshot, so this axis is unscored. It would measure cumulative output volume, venue quality, and the trajectory of papers over time.
Impact—
Insufficient data: citation counts, h-index signals, and downstream uptake are not yet available or meaningful for this 2026-only record, so the axis is unscored. It would measure scholarly influence, adoption, and citations of the work.
Focus0/100
The radar focus score is 0. OpenAlex records five distinct topic tags — speech and dialogue systems (3), topic modeling (2), AI in service interactions (2), plus natural language processing techniques and adversarial robustness (1 each) — while the catalog field split is even (0.5/0.5) between LLM-agent post-training and LLM-agent evaluation. The visible work is thematically linked by banking-domain LLM agents, but the topic spread across adjacent areas keeps the focus reading low.
Collaboration60/100
The collaboration score is 60. The OpenAlex snapshot records 19 distinct co-authors across the recent window, and the frequent-coauthor list is dominated by NPCI colleagues — including Chandra Bhushan, Keyur Doshi, Prashant Devadiga, Yatharth Dedhia and the NPCI AI Research Team as a grouped entry. Both catalog papers are multi-authored team efforts: first-authored IndicBankBench and a middle-position contribution to FiMI Banking, with no corresponding-author role recorded.
Activity40/100
The activity score is 40. The counts-by-year record shows 3 works in 2026 and nothing earlier in the snapshot window, matching recent3yWorks = 3. Output is concentrated in a single year and both catalog items are arXiv preprints, so the cadence is real but short-lived so far; a longer observation window is needed to judge sustained research activity.
Reproducibility60/100
The reproducibility score is 60, averaged over 2 reviewed papers with sharply different readiness. IndicBankBench rates high — a staged benchmark with strict and lenient pass metrics (43.7–58.2% versus 60–74%) — though its judge audit is thin. FiMI Banking rates low: results come from a self-built synthetic environment and self-authored evaluation with no release. Averaging these opposite poles yields the moderate 60.
Research summary
Suvradip Paul is an NLP researcher at the National Payments Corporation of India (NPCI) whose visible work centers on large language model agents for Indian retail banking. An OpenAlex-resolved snapshot (no ORCID linked) records three works in the last three years, all surfacing in 2026, with 19 distinct co-authors — consistent with team-based industrial research. Two catalog-reviewed arXiv preprints define the profile's core. The first-authored IndicBankBench (2026) is a staged benchmark for safety and reliability of language-model assistants in Indian retail banking; Evolve's review credits it with showing that at-least-once success overstates agent reliability (60–74% versus 43.7–58.2% under strict scoring), while noting the LLM-as-judge audit is thin. The co-authored FiMI Banking (2026) trains a 4.5B-parameter sovereign banking agent with DPO and GRPO in a self-built synthetic environment, lifting reward from .610 to .697 and refusal behavior from 52% to 80%; the review flags a self-authored evaluation and no release, so reproducibility readiness is low. OpenAlex field tags split evenly between LLM-agent post-training and LLM-agent evaluation, and topics cluster on speech and dialogue systems, topic modeling, and AI in service interactions. Citation-based impact is not yet measurable — h-index and i10-index read 0 — making this an early-stage, industrially grounded profile.
Collaboration pattern
Suvradip Paul works in a tightly industrial, team-based pattern: all co-authors in the frequent list are NPCI-affiliated, including the grouped entry 'NPCI AI Research Team' and individuals such as Chandra Bhushan, Keyur Doshi, Prashant Devadiga and Yatharth Dedhia. The OpenAlex snapshot counts 19 distinct co-authors in the recent window. On cataloged papers he takes a first-author role on the benchmark (IndicBankBench) and a middle position on the post-training work (FiMI Banking); no corresponding-author role is recorded.
Research keywords
Highlights
IndicBankBench: exposing how lenient scoring overstates banking-agent reliability
FiMI Banking: DPO+GRPO post-training for a sovereign 4.5B banking agent
Timeline
- 2026Publishes IndicBankBench on arXiv as first author — a staged benchmark for safety and reliability of language model assistants in Indian retail banking.
- 2026Co-authors FiMI Banking, post-training a 4.5B sovereign banking agent with DPO and GRPO in a self-built synthetic environment.
- 2026OpenAlex snapshot records 3 works in the recent three-year window with 19 distinct co-authors, split evenly between LLM-agent post-training and LLM-agent evaluation.
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.
- All bibliometrics come from OpenAlex only; no ORCID is linked, so identity is name-resolved from public bibliographic data.
- Citation-based impact is not yet measurable: h-index and i10-index both read 0, and the output and impact radar axes are marked as lacking sufficient data.
- The profile rests on 2 catalog-reviewed arXiv preprints, both from 2026; no peer-reviewed venue record is in scope.
- The reproducibility mean (60) averages only 2 papers with opposite readiness (one high, one low), so it is coarse.