Not whether a model is accurate — whether the decision it drives is fair, honest, and holds under load.

Rajveer Singh Pall

Focus ML & Decision-Systems Research

Institution GGITS, Jabalpur
B.Tech CS & Business Systems

DOWNLOAD RESUME

VIEW RESEARCH

OPTIMISING: DECISIONS · NOT: LEADERBOARDS

Scroll

B.Tech CSBS CS & Business Systems · GGITS Jabalpur
9 Research Papers 6 sole-authored · IEEE, JREFE, JHE +
DS Internships Unified Mentor 2025 · InternPe 2024
IIT Guwahati Data Science Certificate · 2026
SGPA 8.27 Semester V · upward trajectory

01 — About

Accuracy is easy. Fairness under deployment isn't.

I study the systems that decide human outcomes — and whether they deserve the trust we place in them.

Decisions as systems

A mortgage model decides who builds wealth. A clinical model decides who gets treated. A language model decides whose words are believed. My research treats each as a decision system and asks who benefits, who is harmed, and where it fails — then proves the answer with causal evidence at scale.

Honest where it matters

External validation over internal accuracy. Statistical gates that refuse to deploy when the signal is noise. Fairness measured at runtime, not just on a benchmark. Across lending, medicine, language, and markets, the method is the same: triangulate, stress-test, and report where the edge is thin.

What I don't do

No false alpha dressed up as a result. No one-time fairness number that breaks off-distribution. No privacy claims that quietly route data to a cloud API. If a system can't hold under deployment, I'd rather show why than ship it.

Academic record SGPA across five semesters · B.Tech CSBS, GGITS

6.96Sem I
7.41Sem II
7.64Sem III
7.88Sem IV
8.27Sem V

• The Gold and the Glory •

ARIA Assistant

A fully local voice assistant — faster-whisper → Qwen3-8B → five-axis fairness audit → Piper TTS. 16 tools, ≤7.1 GB VRAM on one RTX 4060, zero cloud LLM calls.

Local-firstVoice AIQwen3-8BFairness audit

System2025↗

Comprehensive Evaluation of Machine Learning for Type 2 Diabetes Risk Prediction

XGBoost leads at 0.794 AUC internally but falls to 0.717 on an external population 83× larger — a 9.7% deployment gap. Age and BMI dominate SHAP; race/ethnicity ranks 4th, requiring explicit fairness audit.

XGBoostSHAPFairnessExternal validation

IEEEAccepted · 2025↗

FinSight

Earnings-intelligence pipeline over 14,584 transcripts from 601 S&P 500 firms — FinBERT sentiment fused with RAG-retrieved features, surfacing an Energy-sector information coefficient of +0.31.

FinBERTRAGPipelineEquities

System2025↗

SereneSpace

Anonymous, no-login student mental-health hub with mood-based entry and built-in crisis escalation — selected among the top 50 of 250 teams at Smart India Hackathon 2025.

AnonymousCrisis-awareFlaskWeb

SystemSIH 2025 · 50 of 250↗

When the Gate Stays Closed: Near-Zero Cross-Sectional Predictability in Large-Cap NASDAQ Equities

An IC-gated deployment framework with a conjunctive HAC + permutation test — the gate never opened across 12 folds and 1,512 OOS days, cutting false deployment from 11.8% to 0.0%.

Quant MLWalk-forwardCatBoostPermutation tests

Quantitative Finance & EconomicsUnder review · 2025↗

FinSight + Web

A scroll-driven WebGL monograph that turns the FinSight findings into something a non-technical reader can explore — Next.js 16 + React Three Fiber, shipped on Vercel.

Next.jsReact Three FiberWebGLVercel

System2025–26↗

Privacy-Preserving Federated Learning for Diabetes Risk Across Demographically Heterogeneous Nodes

FedAvg / FedProx / FedNova / SCAFFOLD on partitioned NHANES, externally validated on 1.28M BRFSS records — a 40% smaller generalisation gap than a matched centralised model.

Federated learningPyTorchFlowerDifferential privacy

J. Biomedical InformaticsUnder review · 2026↗

Cross-Platform Generalisation Failure in Mental-Health NLP: A Five-Axis Fairness Audit

Transformer classifiers that score AUC 0.98 within-platform collapse 30–39% off-platform, with calibration and equity failing in lockstep — fairness that doesn't transfer.

NLPFairnessTransformersCalibration

IEEE TNNLSSubmitted · 2026↗

Who Bears the Burden? Heterogeneous Racial Approval Differentials in U.S. Mortgage Lending

Causal Forest Double Machine Learning on 42.3M HMDA applications, 2020–2024 — a 9.4 pp Black approval penalty net of 33 controls, largest under manual underwriting.

Causal inferenceEconMLDMLHMDA

J. Real Estate Finance & EconomicsSubmitted · 2026↗

Persistent Racial Disparities in U.S. Mortgage Approval: Evidence from 42 Million Applications, 2020–2024

Five identification strategies — DFL, within-lender FE, RDD, DiD, Manski bounds — all confirm a 14.95 pp Black–White approval gap, ≥44% of which survives maximally adversarial selection assumptions.

Causal inferenceHMDAPartial identificationRDD

J. Housing EconomicsSubmitted · March 2026↗

IndiaFinBench: Evaluating LLM Performance on Indian Financial Regulatory Text

The first benchmark over SEBI and RBI regulation — 406 expert QA items, 192 documents, 12 LLMs scored zero-shot, plus a hybrid-RAG demo lifting Recall@5 to 0.785.

LLM evaluationRAGFAISSBenchmark

Target: EMNLP 2026In preparation↗

ARIA: Runtime 5-Axis Fairness Audit for Free-Form LLM Output

Inline counterfactual disparate-impact and equalized-odds scoring on a model's actual generated text, not a fixed offline benchmark. Catches 33–39% of failures other guardrails miss, in ~1.2s, with zero cloud calls.

LLM fairnessRuntime auditCounterfactual testing

arXiv preprint2026↗

Overcoming the Transaction Cost Trap: Cross-Sectional Conviction Ranking in ML Equity Prediction

Naive thresholds destroy ML equity alpha through turnover. Treating calibrated ensemble probabilities as ordinal conviction and holding only the Top-1 asset each day gives a significant IC of 0.0197 (p=0.034, 1,512 days) and 45.8% annualised return at Sharpe 1.18 — break-even at 24.2 bps, 4.8× institutional friction.

Quant MLConviction rankingWalk-forwardCatBoost

SSRNWorking paper · 2025↗

fairscope

Subgroup-stratified, calibration-aware fairness auditing as a pip-installable library — DeLong CIs for per-subgroup AUC, per-subgroup ECE/MCE, gap significance testing with Bonferroni/BH correction, and the five-axis CPFE protocol. Five modules at v0.3.0, 100% line coverage on the statistical core, CI green on Python 3.9–3.12.

FairnessCalibrationPyPIOpen source

Systemv0.3.0 · 2026↗

02 — Impact

Proof, not promises.

31 Metrics

ML Research & Deployed Systems * Rajveer Singh Pall

An end-to-end research portfolio spanning causal inference, federated learning, fairness audits, NLP benchmarking, and local AI systems. Hover any node to inspect the metric.

31 metrics across causal inference, federated learning, fairness audits, NLP benchmarking, and local AI systems — the proof behind the work.

03 — Featured Work

Shipped at v0.3.0 on PyPI — 100% line coverage on the statistical core, CI green across Python 3.9–3.12, live documentation site.

GitHub ↗ Documentation ↗ PyPI ↗

04 — Research

Nine papers, one question.

Fairness, healthcare, language, and markets — each result validated externally, ranked here by how far each has gone.

Click card to view

Comprehensive Evaluation of Machine Learning for Type 2 Diabetes Risk Prediction

XGBoost leads at 0.794 AUC internally but falls to 0.717 on an external population 83× larger — a 9.7% deployment gap. Elderly patients reach only 0.607, 13.5 pp below young adults. Age and BMI dominate SHAP; race/ethnicity ranks 4th, requiring explicit fairness audit.

XGBoostSHAPFairnessExternal validation

IEEEAccepted · 2025↗

Persistent Racial Disparities in U.S. Mortgage Approval: Evidence from 42 Million Applications, 2020–2024

Five identification strategies — DFL, within-lender fixed effects, RDD at 80% LTV, DiD across the 2022 tightening, and Manski partial-ID bounds — all confirm a 14.95 pp Black–White approval gap of which ≥44% survives maximally adversarial selection assumptions.

Causal inferenceHMDAPartial identificationRDD

J. Housing EconomicsSubmitted · March 2026↗

Cross-Platform Generalisation Failure in Mental-Health NLP: A Five-Axis Fairness Audit

Transformer classifiers that score AUC 0.98 within-platform collapse 30–39% off-platform, with calibration and equity failing in lockstep — fairness that doesn't transfer.

NLPFairnessTransformersCalibration

IEEE TNNLSSubmitted · 2026↗

Who Bears the Burden? Heterogeneous Racial Approval Differentials in U.S. Mortgage Lending

Causal Forest Double Machine Learning on 42.3M HMDA applications, 2020–2024 — a 9.4 pp Black approval penalty net of 33 controls, largest under manual underwriting.

Causal inferenceEconMLDMLHMDA

J. Real Estate Finance & EconomicsSubmitted · 2026↗

Privacy-Preserving Federated Learning for Diabetes Risk Across Demographically Heterogeneous Nodes

FedAvg / FedProx / FedNova / SCAFFOLD on partitioned NHANES, externally validated on 1.28M BRFSS records — a 40% smaller generalisation gap than a matched centralised model.

Federated learningPyTorchFlowerDifferential privacy

J. Biomedical InformaticsUnder review · 2026↗

When the Gate Stays Closed: Near-Zero Cross-Sectional Predictability in Large-Cap NASDAQ Equities

An IC-gated deployment framework with a conjunctive HAC + permutation test — the gate never opened across 12 folds and 1,512 OOS days, cutting false deployment from 11.8% to 0.0%.

Quant MLWalk-forwardCatBoostPermutation tests

Quantitative Finance & EconomicsUnder review · 2025↗

Overcoming the Transaction Cost Trap: Cross-Sectional Conviction Ranking in ML Equity Prediction

Quant MLConviction rankingWalk-forwardCatBoost

SSRNWorking paper · 2025↗

IndiaFinBench: Evaluating LLM Performance on Indian Financial Regulatory Text

The first benchmark over SEBI and RBI regulation — 406 expert QA items, 192 documents, 12 LLMs scored zero-shot, plus a hybrid-RAG demo lifting Recall@5 to 0.785.

LLM evaluationRAGFAISSBenchmark

Target: EMNLP 2026In preparation↗

ARIA: Runtime 5-Axis Fairness Audit for Free-Form LLM Output

LLM fairnessRuntime auditCounterfactual testing

arXiv preprint2026↗

05 — Stack

What I build with.

Tools that appear in at least one paper or deployed system above.

06 — Contact

Hard problems welcome.

Optimising for: decisions that hold · Not: leaderboard wins

Email GitHub HuggingFace LinkedIn Scholar

Jabalpur, India · --:-- local