Evidence before claims · Baselines before complexity · Limitations documented, not hidden.

I like problems where the hard part is not only training a model — it is building a trustworthy path from raw evidence to a defensible decision.

QUESTION → DATA → METHOD → EVALUATE → DECIDE

My work sits across four connected areas:

• Large-scale Data Science — converting multi-million-row data into reusable analytics, risk signals and forecasts.
• Statistical Experimentation — estimating treatment effects with uncertainty, effect size and practical significance.
• ML Reliability — checking data quality, leakage, imbalance, baselines and explainability before trusting model results.
• Applied AI — combining deterministic analytics with retrieval, reranking, orchestration and grounded generation.

Large-Scale Data Science · NLP · Forecasting

Built an end-to-end intelligence platform on the CFPB Consumer Complaint dataset, turning an 8–9 GB raw CSV into reusable analytical layers, NLP routing, forecasting and decision views.

15.95M complaints · 3.57% forecast MAPE · 75.28% product classifier accuracy · 14/14 core tests

Architecture

8–9 GB CSV
   ↓
Chunk + validate
   ↓
Parquet analytical layers
   ↓
Analytics · NLP · Forecasting
   ↓
Decision views

Key decision — pre-aggregate expensive analysis and reuse validated Parquet outputs instead of loading the multi-GB raw CSV inside the application.

Boundary — complaint volume is not normalized by company customer base; forecasts and risk labels support investigation, not causal claims.

Stack — Python · Pandas · NumPy · PyArrow · Parquet · scikit-learn · TF-IDF · Logistic Regression · Prophet · Streamlit · Docker

Repository · Walkthrough · Evaluation

Statistics · Experimentation · Business Decision-Making

Analyzed a controlled advertising experiment to determine whether the advertisement treatment produced a meaningful conversion improvement over a PSA control.

588,101 users · 2.5547% vs 1.7854% conversion · +0.7692 pp uplift · +43.09% relative uplift

95% uplift interval — +0.5951 to +0.9434 percentage points

Method

Experiment data
   ↓
Validation + group checks
   ↓
Conversion estimates
   ↓
Two-proportion test
   ↓
Confidence intervals + effect sizes
   ↓
Simulation + logistic consistency
   ↓
Power planning
   ↓
Decision

Key principle — never present only a p-value. Report absolute uplift, relative uplift, uncertainty, effect size and practical significance together.

Boundary — ROI is not claimed because campaign cost, incremental revenue and customer lifetime value are unavailable.

Stack — Python · Pandas · SciPy · Statsmodels · Matplotlib · Logistic Regression · Statistical Inference

Repository

ML Reliability · Human-in-the-Loop · MLOps

Built a deterministic-first pre-training audit workflow that checks whether tabular data is sufficiently reliable for responsible baseline modeling.

Dataset
   ↓
Profiling
   ↓
Data Quality · Leakage · Imbalance
   ↓
Risk Aggregation
   ↓
Human Review Gate
   ↓
Baselines
   ↓
MLflow · SHAP
   ↓
Grounded Report + Q&A

Key design — Python owns profiling, checks, calculations, routing and model evaluation. The LLM is restricted to explanation, report generation and follow-up Q&A.

Engineering proof — human review gate · baseline comparison · MLflow tracking · SHAP evidence · FastAPI · pytest · Docker · LangGraph

Boundary — this is not AutoML, a governance certification platform or a replacement for domain review.

Stack — Python · scikit-learn · LangGraph · MLflow · SHAP · FastAPI · Streamlit · pytest · Ruff · Docker

Repository · Live App · Walkthrough

Hybrid Retrieval · Exact Analytics · Grounded AI

Built a document-intelligence system that separates exact structured analytics from semantic retrieval, instead of forcing every question through one RAG path.

CSV / Excel
    ↓
Query Router
    ↓
Pandas Analytics
    ↓
Exact Result

Documents
    ↓
BGE + BM25
    ↓
RRF + Dedup
    ↓
CrossEncoder
    ↓
Grounded Answer + Citations

Strict hybrid RAGAS — 0.947 composite · 0.966 faithfulness · 1.000 context precision · 1.000 context recall

A separate 1,642-case production benchmark passed its defined factual, source, safety, conversation and HTTP checks. Final benchmark acceptance remains intentionally open because the latency gate has not yet passed.

Key decision — deterministic Pandas analytics answer structured questions; hybrid retrieval + reranking handles semantic questions.

Stack — Python · FastAPI · React · Vite · BGE · BM25 · RRF · CrossEncoder · ChromaDB · Pandas · PostgreSQL · Docker · RAGAS

Repository · Live Demo · Walkthrough · Evaluation

01 · Baseline before complexity
Start with the simplest defensible method. Add complexity only when measurement justifies it.

02 · Evaluation before claims
Use holdouts, uncertainty, baselines, failure analysis and explicit acceptance criteria.

03 · Deterministic code before LLM judgment
Calculations, validation and business rules belong in code. LLMs are useful when language understanding or explanation genuinely adds value.

04 · Boundaries are part of the result
A strong project states what was measured, what remains uncertain and what the system cannot prove.

The flagship projects above are portfolio systems. The work below is ongoing practice — kept separate so exercises are not inflated into project claims.

LeetCode public language counts

483 Python3 · 75 MS SQL Server · 36 Pandas

100 Days Badge 2026

View LeetCode profile →

I am pursuing B.Tech Computer Science Engineering (AI & ML) at Uttaranchal University, Dehradun.

My strongest work sits where Data Science meets engineering: large-scale analytics, statistical experimentation, ML reliability, NLP/retrieval and grounded AI systems.

I care about the part after “the model works” — validation, failure modes, reproducibility, interfaces, testing and clearly explaining what the result does not prove.