Aman Nandhra

Overview

An AI-driven Chrome extension that automates resume–job matching. It evaluates how well a resume fits a job description directly in the browser, using semantic understanding, structured requirement extraction, and evidence-based scoring rather than keyword counts.

Methodology

• Designed a modular scoring pipeline that combines structured similarity signals with LLM outputs to rank job fit and provide repeatable recommendations.
• Scores each posting on three parts — eligibility (minimum requirements), relevance (how closely experience aligns) and evidence (proof of similar work) — and combines them into an overall match score.
• Runs on job boards such as HiringCafe, parsing job descriptions automatically and showing results in the page.

Results

• A working, user-facing extension that integrates NLP, similarity scoring, and LLM prompt workflows.
• A feedback loop from marked applications evaluates outcomes and iteratively refines recommendation quality, creating a foundation for continuous model and prompt improvement.

Overview

An end-to-end machine learning workflow forecasting hourly traffic volume on westbound I-94 near Minneapolis–St. Paul, using about 48,000 hourly observations from 2012 to 2018, to forecast demand and analyze system behavior under changing load.

Methodology

• Cleaned the data: duplicate timestamps, a 10-month sensor gap, impossible temperature and rainfall readings, and a road-resurfacing anomaly.
• Engineered temporal, weather, and traffic lag features (1 hour, 24 hours, 1 week).
• Benchmarked Ridge Regression, Poisson GLM, Random Forest, XGBoost, and a stacked ensemble against lookup-table and persistence baselines, using MAE, RMSE, and time-series cross-validation.
• Enforced forecast-issue-time cut-offs at every fold so no model trains on outcomes it could not have seen.

Results

• One-hour-ahead model: MAE 153 vehicles/hour, R² 0.986 — about half the error of the hour-by-weekday lookup table (MAE 300).
• Rolling 24-hour-ahead model: MAE 253 vehicles/hour, R² 0.945.
• Time of day and recent traffic history were the strongest predictors; weather added under 2%, so the models don’t depend on a weather forecast.
• Turned the experimentation into a repeatable forecasting workflow.

Overview

An end-to-end analytics pipeline (SQL → Python → Power BI) over 91,250 records across 50 SKUs, 5 warehouses, and 10 suppliers, analyzing inventory, demand, supplier lead times, and replenishment risk.

Methodology

• Audited the raw data, loaded it into a normalized SQLite database, and cleaned and validated it with quality checks throughout the workflow.
• Applied statistical segmentation and K-Means clustering to group SKUs by value and throughput.
• Evaluated cost–service tradeoffs between 95% and 99% availability targets and summarized the findings in a stakeholder dashboard.

Results

• Identified $523K in persistent excess inventory above the reorder point.
• Quantified the cost of raising service levels from 95% to 99% availability.
• Traced the overstock to static, conservative replenishment parameters.

Overview

An analysis of how player scoring and efficiency change from the regular season to the playoffs, where defensive intensity and game planning rise sharply.

Methodology

• Engineered performance deltas for scoring (PPG) and efficiency (TS%), plus an impact score combining both.
• Grouped players into four quadrants: true improvement, efficient role shift, inefficient volume, and decline.
• Built Python-based visualizations and a multi-page interactive Power BI dashboard (with DAX measures) to surface player-level patterns.

Results

• Over 57% of players saw efficiency decline in the playoffs.
• Across 213 players, scoring fell by about 1.6 points per game and true shooting by 7.7% on average; only 33 improved in both scoring and efficiency.