Arushi Sagar
Pre-final year student · CSE (AI/ML)
AI/ML
Data Analytics
AI/ML, data analytics, and occasionally disappearing down research rabbit holes.
Open to internships, freelance work, collaborations, and interesting problems in general. Reach out if you're hiring or you'd like to work together, or just talk about a research paper or a weird dataset.
a segment from InfraUrban's heat-risk grid — 24 zones × 6 cities, simulated here
01
Projects
PipelineGuardian — LLM-Orchestrated ML Pipeline Auditor
ongoing
Catches data leakage and reproducibility bugs in ML notebooks automatically, calling an LLM only when static analysis alone can't judge the call.
- AST-based static analyzer detecting data leakage and reproducibility gaps across .ipynb/.py files — 1.0 precision/recall on a 12-fixture eval suite, backed by 29 automated tests
- LangChain agent that conditionally invokes an LLM (Groq/Llama-3.3-70B), calling it only when schema signals warrant judgment on train/test split validity — verified against true-negative cases to avoid over-flagging
- Shipped as a CLI, FastAPI backend, and frontend, deployed end-to-end on Render and Vercel
PythonLangChainGroqFastAPIPandasPydanticpytestAST
InfraUrban — Urban Heat Intelligence Platform
Real-time urban heat risk scoring across 24 zones in 6 Indian cities, cutting RMSE 61.9% over baseline with XGBoost and SHAP.
- XGBoost heat-risk model trained on 1M+ hourly climate records across 24 zones in 6 Indian cities
- 61.9% RMSE reduction over a lag-1 baseline, validated on a forward time split to rule out leakage
- SHAP-based per-zone explainability layer plus a custom 5-year warming-slope detector for early risk flagging
- Automated daily-refresh Streamlit dashboard with live geospatial visualization via Folium
PythonXGBoostSHAPStreamlitFolium
Transformer from Scratch — Neural Machine Translation
ongoing
Paper-accurate "Attention Is All You Need" implementation built entirely from raw PyTorch tensors — no HuggingFace shortcuts.
- Full encoder-decoder built from raw PyTorch tensors — custom multi-head attention, positional encoding, and masking, no built-in transformer modules
- Beam search decoding, evaluated with SacreBLEU on the Multi30k German-English benchmark
- Ablation studies isolating the contribution of individual architecture components
- Attention-visualization notebooks showing what each head actually learned
PyTorchNumPySacreBLEU
CycleSphere — Menstrual Health Analytics
Menstrual health analytics platform clustering cycle and lifestyle data, with an interactive what-if simulator.
- KMeans clustering, silhouette-optimized at k=2, on real cycle and lifestyle data
- Kruskal-Wallis testing returned a null result across all three lifestyle factors — reported as-is instead of tuned to produce a positive finding
- Full ETL pipeline plus an interactive what-if simulator, deployed on Streamlit
PythonSQLKMeansStreamlit
Trader Sentiment Analysis
Segments Hyperliquid trader behavior against Bitcoin Fear & Greed regimes to expose distinct strategy archetypes.
- Analyzed Hyperliquid trader behavior against Bitcoin Fear & Greed Index regimes
- Traced a catastrophic timestamp-parsing bug back to major upstream data loss, then rebuilt the ingestion pipeline from scratch
- K-Means (k=3) segmentation separating contrarian from momentum-driven strategies
PythonPandasK-Means
More on GitHub
Smaller experiments, half-finished rabbit holes, and everything not featured here
02
Skills & Tech Stack
Languages
Python · SQL · C++ · C · TypeScript
ML & Deep Learning
Scikit-learn · XGBoost · PyTorch · Transformer Architecture · NLP · Beam Search · Model Evaluation · Feature Engineering
LLM / Agentic Systems
LangChain · Groq · Claude · Prompt-conditioned Agent Design · FastAPI · Pydantic
Data Analytics & Visualization
Pandas · NumPy · SHAP · Plotly · Power BI · Tableau · Folium
Databases
MySQL · SQLite · MongoDB
Tools, Testing & Cloud
Git · GitHub · pytest · Streamlit · Render · Vercel · AWS · VS Code
03
Education
KIET Group of Institutions — B.Tech, Computer Science & Engineering (AI/ML) · 2024–2028
8.27
CGPA / 4 SEMESTERS
100+
LEETCODE SOLVED
3
AWS CERTIFICATIONS
04
Outside the Classroom
Certifications
AWS Certified Cloud Practitioner
AWS Certified Data Engineer – Associate
AWS Certified Machine Learning – Associate
Open Source
merged
Added include/exclude sitemap filtering, merged after a couple rounds of maintainer review.
First OSS contribution — small in scope, but a solid intro to real review cycles.
05
Contact
Got an opportunity, an interesting dataset, or just want to talk AI, analytics, or research papers? I'd love to hear from you.