R
RAJVEER SINGH PALL

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 RESEARCH PROFILE
OPTIMISING: DECISIONS · NOT: LEADERBOARDS
Scroll
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. Leaderboard orderings checked by a human reader, not taken on trust. Fairness measured at runtime, not just on a benchmark. Across lending, medicine, language and network security, 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
  1. 6.96Sem I
  2. 7.41Sem II
  3. 7.64Sem III
  4. 7.88Sem IV
  5. 8.27Sem V
• The Gold and the Glory •
02 — Impact

Proof, not promises.

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

03 — Featured Work

Systems that had to hold.

Deployed software — not notebooks. Each one carries a research result into something you can open and use.

ARIA Assistant

Local-first voice AI · runtime fairness audit

Shipped · Local

Context

Offline fairness numbers mean little if the deployed system drifts. A personal AI should be private, capable, and accountable at runtime.

Approach

A fully local pipeline — faster-whisper → Qwen3-8B → a five-axis audit → Piper TTS — with 16 real tools and zero cloud LLM calls.

System

Runs on a single RTX 4060 (≤7.1 GB VRAM). Every response is scored for calibration, faithfulness, consistency, equity, and attribution before it is surfaced.

Outcome

The equity axis catches 33–39% of failures other guardrails miss — fairness as a live property, on local hardware.

Read the case ↗ GitHub ↗ Live demo ↗

FinSight + Web

Earnings-intelligence pipeline · analytics dashboard

Shipped · Vercel

Context

Can language in earnings calls predict forward returns — honestly, under walk-forward discipline that can't peek at the future?

Approach

14,584 transcripts from 601 S&P 500 firms → FinBERT sentiment + RAG-retrieved features → strict walk-forward backtest.

System

A staged ML pipeline plus a deployed Next.js dashboard rendering the signal field, sector breakdowns, and equity curves.

Outcome

A genuine cross-sectional signal in Energy (IC +0.31) — and an honest report of where the edge is thin elsewhere.

Read the case ↗ Read the web case ↗ Live dashboard ↗ GitHub ↗

IndiaFinBench

LLM evaluation benchmark · Indian financial regulation

Live · HuggingFace

Context

No public benchmark measured how well LLMs read Indian financial regulatory text — SEBI and RBI rulings spanning 1992–2026.

Approach

406 expert-annotated QA items across 192 documents, four reasoning tasks, three difficulty tiers, with a hybrid-RAG open-book demo.

System

12 LLMs scored zero-shot closed-book; a Flask + FAISS + BM25 retrieval demo (RRF) deployed as a HuggingFace Space.

Outcome

The scoring rule decides the leaderboard: strict marking puts the field at 75.12–89.66%, an LLM judge at 87.19–94.83%, and the two orderings barely agree (ρ = −0.273). Hybrid retrieval lifts Recall@5 to 0.785, 9.7 pp over dense alone.

Read the case ↗ Live demo ↗ GitHub ↗

SereneSpace

Anonymous student mental-health platform

SIH 2025 · 50 of 250

Context

The barrier to first contact in student mental health is identity. A login is often enough to stop someone reaching out.

Approach

A no-login entry point routing students to mood check-ins, an AI companion, activities, and peer community.

System

A single web app with crisis-escalation pathways and aggregate, de-identified analytics for administrators.

Outcome

Selected among the top 50 of 250 teams at Smart India Hackathon 2025.

Read the case ↗ Live site ↗

fairscope

Fairness-auditing Python library · PyPI

v0.3.0 · CI green

Context

Mainstream fairness toolkits don't expose subgroup-stratified statistical machinery — DeLong intervals, per-subgroup calibration, corrected gap tests — as first-class functions.

Approach

Package peer-reviewed methods behind one-call domain audits (healthcare, lending, federated), each citing its source; one novel protocol on top — the five-axis CPFE.

System

Five modules on NumPy/SciPy/sklearn with optional NLP and lending extras; reports with forest plots, reliability diagrams, and pre/post-recalibration ECE.

Outcome

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.

Read the case ↗ GitHub ↗ Documentation ↗ PyPI ↗
04 — Research

Six papers, one question.

Fairness, healthcare, LLM evaluation and network security, ordered by their weight in the research program.

01
Click card to view
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 Research Profile

Jabalpur, India · --:-- local