Yatharth Dedhia

National Payments Corporation of India · India

Unclaimed · AI-structured profileOpenAlex

LLM agent reliability and post-training for Indian retail banking — applied AI researcher at NPCI

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—

Not yet scored: OpenAlex citation-indexed coverage for this author is insufficient — h-index and i10-index both stand at 0, with only 4 works in the recent three-year window. This axis would measure publication volume and productivity once the record consolidates.

Impact—

Not yet scored: the snapshot contains no usable influence data — h-index and i10-index are 0 and no citation counts are available. This axis would measure citations, adoption, and downstream use of the benchmarks and models once such data accumulates.

Focus0/100

Focus scores 0 in the snapshot, reflecting topical spread: OpenAlex topics span five clusters — speech and dialogue systems (3), topic modeling (2), AI in service interactions (2), plus single counts in advanced neural network applications and NLP techniques. The field distribution itself is tightly bimodal, splitting 0.5/0.5 between LLM agents/post-training and LLM agent evaluation — two adjacent subfields serving one banking-agent agenda.

Collaboration60/100

Collaboration scores 60, computed from 19 distinct coauthors in the recent three-year window — an industry-team pattern anchored at NPCI. Frequent coauthors include Chandra Bhushan, Keyur Doshi, Prashant Devadiga, Suvradip Paul, and the group-listed NPCI AI Research Team. On the two reviewed papers Dedhia sits middle (IndicBankBench) and last (FiMI Banking), neither corresponding — a contributor or senior-contributor role within a larger group.

Activity47/100

Activity scores 47 on a young record: OpenAlex countsByYear show 1 work in 2025 and 3 in 2026, with 4 works across the last three years. Both reviewed papers landed in 2026, indicating a rapid ramp-up from a standing start rather than an established annual publication cadence.

Reproducibility60/100

Reproducibility scores 60 (reproMean 60 across 2 repro-scored papers). The internal split is instructive: IndicBankBench earns high reproducibility readiness as a staged benchmark, while FiMI Banking rates low because its evaluation is self-authored and no model release accompanies it — openness tracks the evaluation artifact, not the trained model.

Research summary

Yatharth Dedhia is an early-career AI researcher at the National Payments Corporation of India (NPCI) whose recent work concentrates on making large language model agents reliable in a high-stakes, regulated domain: Indian retail banking. OpenAlex attributes 4 works in the last three years (1 in 2025, 3 in 2026) across 19 distinct coauthors, with field distribution split evenly (0.5/0.5) between NLP for LLM agents and post-training and NLP/LLM agent evaluation. Evolve has reviewed two 2026 arXiv preprints. IndicBankBench, a staged safety-and-reliability benchmark for language model banking assistants, shows that at-least-once success (60–74%) overstates true capability relative to strict success (43.7–58.2%); Evolve rates its reproducibility readiness high while flagging a thin judge audit. FiMI Banking, where Dedhia is last author, applies DPO plus GRPO to a 4.5B model inside a self-built synthetic environment, lifting reward from 0.610 to 0.697 and refusal behavior from 52% to 80%, though the evaluation is self-authored and no weights are released. Topic signals (speech and dialogue systems, topic modeling, AI in service interactions) point to a service-industry conversational AI focus. Citation-backed output and impact are not yet computable (h-index and i10-index stand at 0 in the snapshot), marking this as a young, industry-embedded profile best read as a benchmark-plus-training agenda taking shape.

Collaboration pattern

Collaboration (score 60) is industry-team science: 19 distinct coauthors in the recent three-year window, anchored at NPCI. Frequent coauthors include Chandra Bhushan, Keyur Doshi, Prashant Devadiga, Suvradip Paul, and the group-listed NPCI AI Research Team. On the two reviewed papers Dedhia holds middle and last positions, neither corresponding — consistent with a senior contributor inside a larger AI group rather than a solo corresponding lead.

Research keywords

LLM agentsAgent benchmarkingTool useReliability evaluationpass@k metricsLLM-as-judgeDirect preference optimization (DPO)GRPO / reinforcement learning with verifiable rewardsSynthetic data generationDomain-specific LLMsSpeech and dialogue systemsIndian retail banking NLP

Highlights

IndicBankBench: staged benchmark exposes overstated banking-agent reliability
This 2026 benchmark evaluates language model assistants in Indian retail banking and shows that at-least-once success rates of 60–74% fall to 43.7–58.2% under strict success criteria. Evolve calls the staged design rigorous and rates its reproducibility readiness high, while noting that the LLM-as-judge audit remains thin.
FiMI Banking: DPO+GRPO post-training for a sovereign 4.5B banking agent
As last author, Dedhia contributed to a self-built synthetic environment where DPO plus GRPO lift a 4.5B model's reward from 0.610 to 0.697 and refusal behavior from 52% to 80%. Evolve credits a solid method but flags the self-authored evaluation and the absence of any release.
A benchmark-to-training loop for Indian banking AI
Read together, the two 2026 papers form a full-stack agenda: IndicBankBench supplies an external measurement of banking assistants (60–74% at-least-once vs 43.7–58.2% strict) while FiMI Banking trains the model (reward 0.610→0.697, refusal 52%→80%). The open question Evolve flags is closing that loop with independent evaluation and a public release.

Timeline

2025First OpenAlex-indexed work appears: the record begins with a single publication in 2025, marking the start of measurable research activity.
2026Co-authors IndicBankBench (middle position) — a staged safety-and-reliability benchmark for language model assistants in Indian retail banking, posted to arXiv.
2026Co-authors FiMI Banking as last author — DPO+GRPO post-training of a 4.5B sovereign banking agent inside a self-built synthetic environment.
2026Output accelerates: OpenAlex records 3 works in 2026, with 19 distinct coauthors across the recent three-year window.

Papers in the Evolve catalog

Frequent co-authors

Chandra Bhushan×2 Keyur Doshi×2 Prashant Devadiga×2 Suvradip Paul×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.

Bibliometrics are from OpenAlex only; no ORCID is linked, so citation aggregation may undercount an industry author's output.
Output and impact axes are unscored: citation coverage is insufficient in the snapshot, with h-index and i10-index both 0.
Only the two 2026 arXiv preprints listed here have been reviewed by Evolve; their peer-review status is not reflected in this data.
Profile is name-resolved from public bibliographic data; identity confidence depends on the OpenAlex author record for this name.