Nokia researcher: LLM agent evaluation & post-training for Indian retail banking (benchmarks, DPO/GRPO, reliability)
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
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—
This axis measures publication volume and productivity over time. OpenAlex indexes very few records for this identity (h-index 1), so no output score can be computed yet; the five works indexed in 2026 suggest new activity that is still being captured.
Impact—
This axis measures citation influence — citations, h-index and i10-index relative to field norms. With an h-index of 1 and an i10-index of 1, citation data is insufficient to score; the two 2026 preprints are too recent to have accumulated citations.
Focus0/100
The focus score is 0. The field split is exactly even — 50% LLM agents and post-training, 50% LLM agent evaluation — two adjacent subtopics, while OpenAlex topics range across speech and dialogue systems (3), topic modeling (2), AI in service interactions (2) and one nanoparticle drug-delivery entry, indicating breadth across domains rather than a narrow single-line research program.
Collaboration60/100
Scoring 60, this axis reflects 19 distinct coauthors in the recent window and a dense recurring cluster — Chandra Bhushan, Prashant Devadiga, Suvradip Paul, Yatharth Dedhia and an NPCI AI Research Team affiliation. Both cataloged 2026 papers list Doshi in a middle author position rather than first or corresponding, consistent with contribution to a larger applied team effort on banking-domain LLM agents.
Activity52/100
With an activity score of 52, the record shows episodic output: one indexed work in 2017, one in 2022, then a burst of five works in 2026, all within the last three years (recent3yWorks: 5). The 2026 cluster — two arXiv preprints on Indian retail-banking LLM agents — marks a clear re-entry into active research after a multi-year quiet stretch.
Reproducibility60/100
Reproducibility scores 60, computed from two assessed papers (reproMean 60). The picture is split: IndicBankBench rates high on readiness as a staged, transparently described benchmark, while FiMI Banking rates low because its evaluation is self-authored with no model or environment release. This contrast — rigorous external benchmarking versus closed post-training work — defines the current reproducibility profile.
Research summary
Keyur Doshi is a researcher at Nokia in Finland whose recent work centers on evaluating and post-training LLM agents for high-stakes, domain-specific deployments, with an OpenAlex record split exactly 50/50 across two adjacent NLP subfields: LLM agent evaluation and LLM agents with post-training. The Evolve catalog includes two 2026 arXiv preprints. IndicBankBench introduces a staged benchmark for the safety and reliability of language-model assistants in Indian retail banking, showing that at-least-once success (60–74%) substantially overstates reliability versus strict pass rates (43.7–58.2%); Evolve rates its reproducibility readiness high while noting a thin judge audit. FiMI Banking applies DPO and GRPO to a 4.5B-parameter banking agent trained in a self-built synthetic environment, lifting reward from 0.610 to 0.697 and refusal behavior from 52% to 80%; Evolve credits the method but flags the self-authored evaluation and lack of release, rating reproducibility low. The broader OpenAlex footprint is thin and episodic — h-index 1, one indexed work in 2017, one in 2022, and five in 2026 — with topics spanning speech and dialogue systems, topic modeling, and AI in service interactions alongside a single nanoparticle drug-delivery entry. Nineteen distinct coauthors appear in the recent window, anchored by a recurring cluster including an NPCI AI Research Team affiliation, and both cataloged papers are middle-authored rather than first or corresponding — an early-stage, applied, team-based research profile.
Collaboration pattern
Collaboration scores 60, with 19 distinct coauthors in the recent window. Both 2026 papers are middle-authored and non-corresponding, anchored by a dense recurring cluster: Chandra Bhushan, Prashant Devadiga, Suvradip Paul, Yatharth Dedhia and an NPCI AI Research Team affiliation, alongside Aman Kumar, Asit Desai and others. The pattern points to applied, team-based research on banking-domain LLM agents rather than solo-led work.
Research keywords
Highlights
Staged benchmark exposes over-optimistic banking-agent reliability
DPO+GRPO post-training lifts a 4.5B banking agent
Timeline
- 2017First indexed work appears in the OpenAlex record (1 work in 2017).
- 2022One further indexed work (2022), before a multi-year quiet stretch.
- 2026Most active year on record: 5 indexed works, including two arXiv preprints in the Evolve catalog — IndicBankBench and FiMI Banking on Indian retail-banking LLM agents.
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.
- Bibliometrics come from OpenAlex only; the record is thin (h-index 1, i10-index 1), and the output and impact dimensions lack sufficient data to score.
- Both cataloged papers are 2026 arXiv preprints, not yet published in peer-reviewed venues.
- Identity is name-resolved from public bibliographic data; "Keyur Doshi" is a fairly common name, so coverage beyond the cited ORCID/OpenAlex record may include other authors.
- Only 2 papers have been reviewed by Evolve, so the reproducibility score (60) rests on just those two assessments.