Subasish Das

Texas State University

Unclaimed · AI-structured profileOpenAlex

NLP researcher building auditable, calibrated AI for road-safety analytics and social-media mental-health benchmarks

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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 is unscored: the snapshot contains no publication counts, venue history, or per-year output totals for this researcher. It would measure the volume, venue quality, and consistency of published work over the full career record.

Impact—

Impact cannot be scored: no citation counts, citation-based indices, or venue influence data are present in the input. This axis would measure how frequently and widely other researchers cite, reuse, and build on this work.

Focus0/100

The focus score is 0, and the raw field distribution explains why: it splits exactly evenly at 0.33 across three areas — natural language processing for road-safety analytics, applied machine learning for transportation safety, and computational social-media mental-health NLP. This reads as deliberate breadth rather than drift, since all three lines share one methodological core of weak supervision, calibration, and auditability.

Collaboration36/100

The collaboration score is 32, computed from 4 distinct co-authors across the reviewed record: Gaurab Chhetri, Amir Rafe, Anika Baitullah, and Anandi Dutta. Das sits in the last author position on all three reviewed arXiv papers and is corresponding author on none — a compact, recurring team pattern consistent with a senior supervisory role rather than large multi-institution collaboration.

Activity—

Recent activity is unscored: apart from three papers dated 2026, the snapshot provides no year-by-year productivity data. This axis would measure publication cadence and sustained research engagement across the most recent years of the record.

Reproducibility43/100

The reproducibility score is 43, a mean over the 3 reviewed papers, whose Evolve readiness ratings are medium, low, and medium. The audits reward candid label auditing and leakage control, but cite thin human labels on rare variables, mental-health label agreement of κ=0.222, and a single data split — partial reproducibility with clearly identifiable gaps.

Research summary

Subasish Das works at the intersection of natural language processing and applied safety analytics, holding the last author position on all three Evolve-reviewed 2026 arXiv papers — a pattern consistent with a senior or supervisory role on a small team built around four recurring co-authors: Gaurab Chhetri, Amir Rafe, Anika Baitullah, and Anandi Dutta. The research program is unusually coherent: making large-scale NLP systems auditable, calibrated, and honest about their own label noise. One paper converts 195,857 Texas police crash narratives into probabilistic crash variables at a reported cost of $55.79 with explicit human review budgets. Another reframes dementia-related crash severity prediction as selective, deferrable triage with outcome-leakage control. A third builds a validation-aware weak-supervision benchmark for suicidal ideation and mental-health signals on Bluesky, candidly reporting that mental-health labels are noisy (κ=0.222), so downstream scores track label noise. Evolve's verdicts consistently credit this methodological honesty while flagging thin human labels on rare variables, single data splits, and weak absolute performance. The field distribution splits exactly evenly across road-safety NLP, transportation-safety AI, and computational social-media mental-health NLP, which yields a focus score of 0; reproducibility averages 43 across the three papers, while output, impact, and activity metrics are unavailable in this snapshot and left unscored.

Collaboration pattern

Collaboration is compact: only 4 distinct co-authors appear across the three reviewed papers — Gaurab Chhetri, Amir Rafe, Anika Baitullah, and Anandi Dutta — yielding a collaboration score of 32. Das takes the last author position on every reviewed arXiv paper and is corresponding author on none, a pattern consistent with supervising a small, recurring team rather than joining large multi-institution consortia.

Research keywords

crash narrativesprobability calibrationselective predictionLLM evaluationhuman-in-the-loop reviewweak supervisionsuicidal ideation detectionmental health NLPcrash severity predictionpolice narrative miningoutcome leakage controlBluesky

Highlights

Coding 195,857 crash narratives into calibrated probabilities for $55.79
Evolve's audit of this System One model paper credits strong methods: a typed decision model converts Texas police crash narratives into probabilistic crash variables with explicit human review budgets, while human labels remain thin on rare variables. Reproducibility readiness rated medium.
Candid label audits for a Bluesky mental-health benchmark
This validation-aware weak-supervision framework builds suicidal-ideation and mental-health label sets on Bluesky. Evolve finds suicidal-ideation labels validate fairly well but mental-health labels are noisy (κ=0.222), so mental-health scores track label noise; readiness rated low.
Auditable, deferrable AI triage for dementia-related crash severity
This paper reframes crash severity prediction as auditable AI triage with selective deferral to human review and outcome-leakage control. Evolve credits honest statistics but flags a single data split, weak absolute performance, and unevaluated deferred cases; readiness rated medium.

Timeline

2026Published (arXiv, last author) a System One model converting Texas police crash narratives into probabilistic crash variables with explicit review budgets.
2026Published (arXiv, last author) a validation-aware weak-supervision benchmark for suicidal ideation and mental-health signals on Bluesky, reporting mental-health label agreement of κ=0.222.
2026Published (arXiv, last author) a selective deferral framework for dementia-related crash severity prediction with outcome-leakage control to support human review.

Papers in the Evolve catalog

Frequent co-authors

Gaurab Chhetri×2 Amir Rafe×1 Anika Baitullah×1 Anandi Dutta×1 Subasish Das×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 scores are computed from a public bibliographic snapshot; no institution, country, or ORCID was provided, so this profile is name-resolved and same-name researchers may be merged or missed.
Output, impact, and activity dimensions are unscored because the input contains no citation counts, publication totals, or per-year productivity data.
Scope covers only three Evolve-reviewed 2026 arXiv preprints; peer-reviewed venues, earlier publications, and the full career record are not represented.
The reproducibility score (43) averages only 3 papers, so a single paper's readiness rating can move it substantially.