Prashant Devadiga

National Payments Corporation of India · India

Unclaimed · AI-structured profileOpenAlex

LLM agent post-training and evaluation for Indian retail banking at NPCI — builder of the IndicBankBench benchmark

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Output metrics are not yet scored: the record shows only 5 recent-window works and an h-index of 0, too thin a base to characterize productivity. This axis would measure total and per-year publication volume once a longer record exists.

Impact—

Impact is not yet scored: the snapshot shows an h-index and i10-index of 0 with no citation series, and both catalog papers are 2026 preprints. This axis would track citation uptake and scholarly influence as the work ages.

Focus0/100

The focus score is 0, reflecting a deliberately two-sided profile: the field distribution splits evenly at 0.5 between LLM-agent post-training and LLM-agent evaluation, and OpenAlex topics span five areas (topic modeling and speech/dialogue systems at 3 each, NLP techniques and AI in service interactions at 2, information retrieval at 1). The application domain stays narrow — Indian retail banking — even as methodological topics remain broad.

Collaboration60/100

The collaboration score is 60, built on 19 distinct coauthors in the recent window. The frequent-coauthor list is dominated by an NPCI-centered group — Chandra Bhushan, Keyur Doshi, Suvradip Paul, Yatharth Dedhia and the NPCI AI Research Team among twelve recurring names — consistent with team-produced arXiv work. On the two catalog papers the researcher holds a last authorship on IndicBankBench and a middle position on FiMI Banking, neither corresponding.

Activity52/100

The activity score is 52. The counts-by-year series shows 5 works in 2026 and nothing recorded before it, with recent3yWorks also at 5 — so visible output is recent and concentrated in a single year rather than sustained across a longer span. Both catalog papers are 2026 arXiv preprints, which supports an early-stage, industrially embedded rhythm whose longer-term cadence cannot yet be confirmed.

Reproducibility60/100

The reproducibility score is 60, a mean over 2 assessed papers with divergent readiness. IndicBankBench rates high — a staged, pass@k-based benchmark whose strict results (43.7–58.2%) are reported against at-least-once figures (60–74%) — while FiMI Banking rates low, with a self-authored evaluation environment and no model release noted by the review. The average therefore mixes a reusable evaluation artifact with a non-released model.

Research summary

Prashant Devadiga is a researcher at the National Payments Corporation of India (NPCI) whose published work to date centers on making language-model assistants dependable in Indian retail banking. The Evolve catalog contains two 2026 arXiv papers spanning the two halves of the agent pipeline: FiMI Banking, a sovereign 4.5B banking agent post-trained with DPO and GRPO in a self-built synthetic environment (reward rising from .610 to .697 and refusal from 52% to 80%), and IndicBankBench, a staged safety and reliability benchmark showing that at-least-once success rates of 60–74% overstate agent reliability relative to strict results of 43.7–58.2%. The field distribution splits evenly (0.5/0.5) between LLM-agent post-training and LLM-agent evaluation, while OpenAlex topics add breadth across topic modeling, speech and dialogue systems, NLP techniques, AI in service interactions, and information retrieval. Bibliometric signals are early-stage: 5 works recorded in 2026, an h-index of 0, and 19 distinct coauthors, with a recurring NPCI-centered author group. Evolve rates reproducibility at a mean of 60 over two papers — high for the benchmark, low for the model — reflecting a profile that builds both evaluation infrastructure and models, with releases and independent judging the main gaps. This profile is name-resolved from public bibliographic data.

Collaboration pattern

Collaboration is team-based and institutionally concentrated: of 19 distinct recent coauthors, the recurring core is NPCI's own AI group — Chandra Bhushan, Keyur Doshi, Suvradip Paul, Yatharth Dedhia and others across twelve named collaborators. The two catalog papers are built on this shared author group, with a last-listed position on IndicBankBench and a middle position on FiMI Banking; no external-academic partners are visible in this snapshot.

Research keywords

LLM agentsLLM agent evaluationBenchmark designMulti-turn tool usepass@k metricsLLM-as-judgeReliability of language model assistantsReinforcement learning with verifiable rewardsGRPODirect preference optimizationSynthetic data generationDomain-specific LLMs for banking

Highlights

At-least-once success overstates banking-agent reliability
In IndicBankBench, language-model assistants in Indian retail banking score 60–74% under at-least-once success but only 43.7–58.2% under strict matching. Evolve calls it a rigorous staged benchmark while noting the judge audit is thin.
A staged safety benchmark for Indian retail banking
IndicBankBench provides a staged evaluation of safety and reliability for LM assistants in Indian retail banking; Evolve rates its reproducibility readiness high, making it the more reusable of the researcher's two catalog artifacts.
DPO+GRPO lifts a 4.5B sovereign banking agent
FiMI Banking post-trains a 4.5B model in a self-built synthetic environment, lifting reward from .610 to .697 and refusal from 52% to 80%; Evolve credits the method but flags the self-authored evaluation and lack of a release.

Timeline

2026Co-authored FiMI Banking on arXiv (middle position): DPO+GRPO post-training of a 4.5B banking agent lifts reward from .610 to .697 and refusal from 52% to 80%.
2026Last-listed author on IndicBankBench, a staged safety and reliability benchmark for language-model assistants in Indian retail banking.
2026OpenAlex records 5 works in the year, matching the recent three-year window total and marking the first visible output year.
2026The collaboration network reaches 19 distinct coauthors, centered on the recurring NPCI AI Research Team author group.

Papers in the Evolve catalog

Frequent co-authors

Chandra Bhushan×2 Keyur Doshi×2 Suvradip Paul×2 Yatharth Dedhia×2 NPCI AI Research Team×1 Aman Kumar×1 Asit Desai×1 Harsh Sharma×1 Harshit Bhushan×1 Hrithik Kadam×1 Kolisetty Sai Kapardheeswar×1 Krishanu Adhikary×1

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.

Bibliometric metrics (h-index, i10-index, counts by year, coauthor counts) come from an OpenAlex identity snapshot; no ORCID is linked, so name resolution relies on public bibliographic data.
Qualitative findings rest on only two 2026 arXiv preprints reviewed by Evolve; no peer-reviewed venue record appears in the snapshot.
Citation-based output and impact scoring is unavailable, and no citation counts beyond the raw h-index and i10-index values of 0 are quoted.
Institutional affiliation is taken from the input record; collaboration patterns beyond the listed coauthors are not verified.