Sourav Nandy

The University of Texas at Austin · United States

Unclaimed · AI-structured profileOpenAlex

UT Austin researcher working across LLM agent security, inference-cost routing, and executable AI governance

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-09) 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 would measure total publication volume and its spread across venues and years. The current OpenAlex snapshot records only four works, all from 2026, which is insufficient to compute a stable output score for this profile.

Impact—

Impact measurement needs accumulated citations over time. The snapshot reports an h-index of 0 and an i10-index of 0 with no citation history, so no impact score can be assigned to this very early-stage record.

Focus0/100

The focus score of 0 reflects a perfectly even field distribution: computer security (LLM agent security / prompt injection), AI governance / computational law, and LLM inference optimization each account for exactly one third of the record. Five distinct topics carry one work each — from topic modeling to AI ethics — so output is deliberately broad across three complementary areas rather than concentrated in a single specialty.

Collaboration14/100

The collaboration score of 14 matches a very compact network: exactly 1 distinct coauthor appears in the snapshot, and the sole frequent partner is Rudrendu Kumar Paul. All three Evolve-catalogued 2026 preprints place Nandy in the last author position without corresponding-author designation, a pattern consistent with a senior contribution within a small two-person team rather than participation in large multi-institution groups.

Activity47/100

Activity scores 47, driven entirely by a 2026 burst: the counts-by-year data shows 4 works in 2026 and none in earlier years, with recent3yWorks also at 4. All three reviewed preprints appeared on arXiv in 2026. This concentrated recent output indicates an active current phase, but the absence of any earlier record means cadence and stamina over multiple years cannot yet be judged.

Reproducibility20/100

The reproducibility score of 20 is the mean across 3 reviewed papers, all rated low repro-readiness by Evolve. Specific limits named in the reviews: AgentRouter reports a 71.8% cost cut with no statistical significance tests; the multi-agent injection defense study evaluates a single system without confidence intervals; and Governance-as-Code was validated on two private case studies with no code release.

Research summary

Sourav Nandy is a researcher at The University of Texas at Austin whose visible scholarly record, resolved through OpenAlex bibliographic data, consists of 2026 arXiv preprints at the intersection of LLM systems engineering, agent security, and AI governance. Evolve has reviewed three of these works, on each of which Nandy is the last-listed author. AgentRouter proposes heterogeneous model routing for cost-optimal multi-step agentic workflows, claiming a 71.8% inference cost reduction while preserving 97.3% quality. A second preprint formalizes a threat model of 14 prompt-injection vectors in multi-agent LLM systems and a defense architecture of 4 mechanisms that cuts attack success from 31.2% to 4.2%. The third, Governance-as-Code, translates EU AI Act Articles 8-15 into 43 executable Rego policy checks integrated with CI/CD deployment gates. Bibliometric signals are early-stage: 4 works recorded in 2026 with none earlier, an h-index and i10-index of 0, a single distinct coauthor (Rudrendu Kumar Paul), and low reproducibility readiness across all reviewed papers. Evolve's verdicts are consistently promising but methodologically cautious, citing oracle fallbacks, single-system evaluations, missing significance tests, and no code release. The overall pattern is a focused, output-active early profile whose three-pronged agenda — cheaper agents, safer agent interactions, and auditable compliance — is coherent even though citation impact has not yet accumulated.

Collaboration pattern

Nandy's recorded network is minimal: exactly 1 distinct coauthor, with Rudrendu Kumar Paul as the only frequent partner, yielding a collaboration score of 14. On all three catalogued 2026 arXiv preprints Nandy sits in the last author position without corresponding-author status, consistent with a senior-contributor role in a small two-person team. No multi-institution or large-consortium collaborations appear in the available data.

Research keywords

LLM routingAgentic workflowsInference cost optimizationModel cascadesPrompt injectionMulti-agent LLM systemsThreat modelingDefense architectureEU AI ActGovernance-as-CodePolicy-as-code (Rego/OPA)LLM robustness testing

Highlights

AgentRouter: step-level routing with a claimed 71.8% cost cut at 97.3% quality
The 2026 preprint routes each step of multi-step agentic workflows across heterogeneous models, reporting a 71.8% inference cost reduction while preserving 97.3% quality. Evolve calls it promising and well-baselined, but flags reliance on an oracle fallback and the absence of significance tests.
14-vector threat model for prompt injection in multi-agent LLM systems
The paper formalizes 14 injection vectors spanning multi-agent pipelines and proposes a defense architecture of 4 mechanisms, including message signing, that cuts attack success from 31.2% to 4.2%. Evolve values the taxonomy but notes the single-system evaluation and missing confidence intervals.
EU AI Act Articles 8-15 compiled into 43 executable Rego checks
Governance-as-Code translates the EU AI Act's technical requirements into 43 executable Open Policy Agent / Rego checks wired into CI/CD deployment gates. Evolve rates the design strong and legally careful, but it was validated only on two private case studies with no code released.

Timeline

2026Posts the AgentRouter preprint on heterogeneous model routing for cost-optimal multi-step agentic workflows, listed as last author.
2026Publishes a threat model of 14 prompt-injection vectors in multi-agent LLM systems with a 4-mechanism defense architecture, as last author.
2026Releases Governance-as-Code, translating EU AI Act Articles 8-15 into 43 executable Rego checks with CI/CD deployment gates, as last author.
2026OpenAlex records 4 works in 2026 — the entire visible output of the recent three-year window — with the h-index and i10-index still at 0.

Papers in the Evolve catalog

Frequent co-authors

Rudrendu Kumar Paul×3

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 computed from an OpenAlex snapshot only; no ORCID is linked, so identity resolution relies on public bibliographic data.
Citation impact is effectively unmeasured: h-index 0 and i10-index 0, so any impact-related judgement is premature.
Qualitative assessments rest on three Evolve-reviewed 2026 arXiv preprints; no peer-reviewed venue record is in scope.
All three papers carry low repro-readiness ratings, and Evolve verdicts are review judgements rather than independent replications.