$ whoami

Nilesh Singh —
I build systems that measure themselves.

BCA Data Science student in Pune. I build end-to-end analytics and ML systems — and when I claim something works, I go build the test that checks whether it actually does. Data Analyst · ML Engineer · AI Trainer.

About

I'm a final-year BCA student specializing in Data Science at Sri Balaji University, currently completing a Cloud Application Developer internship (AWS, SQL, PySpark). Outside of coursework I build full projects — cleaning the data, training the models, and shipping something a non-technical person could actually use.

My most recent project is a controlled benchmark for evaluating LLM safety across nine dimensions. The part I'm proudest of isn't the benchmark itself — it's that I didn't just trust my own scoring system. I hand-labeled real model responses, measured how well an AI judge agreed with my own judgment, and reported the honest result even when it wasn't flattering.

4
shipped projects
9
eval dimensions benchmarked
7.0
CGPA
2026
graduating

// composite benchmark scores — llm-safety-eval-benchmark

scale 1–5
GPT-OSS-120B
4.15
GPT-OSS-20B
4.14
Qwen3.6-27B
3.83
judge validated via calibration (11/11 on reference cases) + human-agreement study (Pearson r=0.58, quadratic weighted κ=0.31, n=12) — full methodology in the repo →

Selected work

LLM Safety & Response Evaluation Benchmark FEATURED

github ↗ live demo ↗

A controlled benchmark evaluating AI responses across 9 dimensions — instruction following, factuality, relevance, bias, toxicity, refusal quality, prompt injection resistance, hallucination, and consistency — run entirely on free-tier APIs. Includes a 2-model LLM-judge ensemble, a judge calibration test, and a human-agreement study measured with Pearson/Spearman correlation and quadratic weighted kappa. Mid-project, diagnosed and recovered from a live Groq model deprecation without invalidating prior results.

PythonGroq APILLM-as-judge pandasmatplotlibJupyter

Customer Churn Analysis

github ↗

A production-style churn analysis on a 7,000+ row Telco dataset: SQL views, data-quality checks, cross-validation, cost-sensitive thresholds, and a human-review dashboard. The balanced model was selected on business trade-offs, not just accuracy.

Pythonpandasscikit-learn Streamlit

AI-Powered HR Automation Suite

github ↗

Six automated workflows (n8n + GPT-4) covering onboarding, leave management, sentiment analysis, policy Q&A, resume screening, and a WhatsApp HR assistant — including a PDF-parsing resume screener with automated Slack alerts and Google Sheets logging.

n8nGPT-4Google Sheets Slack/WhatsApp

LinguaQ

github ↗

A Streamlit web app that detects the language of input text in real time — built to ship something small and usable end-to-end, not just live in a notebook.

PythonStreamlitlangdetect

Skills

Programming & ML

  • Python, SQL, R
  • Classification, Decision Trees, Clustering
  • Deep Learning fundamentals
  • Model evaluation

LLM Evaluation & AI

  • Prompt engineering (GPT-4)
  • LLM-as-judge design
  • Judge calibration & human-agreement analysis
  • n8n workflow automation

Analytics & BI

  • Power BI, Excel (Solver)
  • Exploratory data analysis
  • Statistical analysis
  • Dashboarding

Tools & Platforms

  • Jupyter, Streamlit
  • Git / GitHub
  • AWS, PySpark
  • Data structures & algorithms