Overview

HR Trends is a platform I built during my internship at Intrizen that turns public U.S. labor-market data into something a business can actually use day to day — tracking indicators like the unemployment rate, employment rate, and education-to-job ratio — instead of leaving that data sitting in static government reports.

The Problem

Government labor data from the Bureau of Labor Statistics (BLS) is public, current, and genuinely valuable — but it isn't built to be used. It's spread across static reports on its own update schedule, with no easy way to turn it into a live view of what's actually changing in the labor market. Intrizen needed a way to make that data continuously useful instead of a one-time pull.

Separately, on the marketing side, I found a version of the same problem on Intrizen's own LinkedIn page: good raw material — roughly 30 employees — that wasn't turning into visibility. Only 6 employees showed up as connected to the company page, and organic engagement was stuck around 30 people a month.

My Role

I was the sole Project & Marketing Intern on this, which meant I owned it end to end: the data pipeline, the website and admin dashboard, the AI layer on top of it, and the brand identity around it — plus the separate analysis of the company's LinkedIn performance.

Process

The pipeline pulls data from BLS, runs it through automated n8n workflows, and stores it in Supabase, which powers both the public website and an internal admin dashboard. On top of that, I built an agentic AI workflow that can generate content, visuals, and insight summaries directly from the underlying data.

What I Built

• Automated data pipeline (BLS → n8n → Supabase)
• Public website and internal admin dashboard
• Agentic AI workflow for generating content, visuals, and insight summaries
• Brand identity — logo and a sea-green visual identity for the platform
• A LinkedIn growth strategy proposal (employee advocacy + AI-generated content), based on the engagement analysis above

Outcome

HR Trends shipped as a live platform that Intrizen uses internally to pull labor-market insight into its marketing and business work. The LinkedIn analysis and strategy were delivered as recommendations — I don't have post-adoption engagement numbers to report yet, so I'm not listing any.

Reflection

Overview

As part of Fair Lawn's municipal Green Team advisory committee, our team calculated the Borough's Scope 2 greenhouse-gas emissions — electricity and natural gas usage across 93 municipal accounts, sourced directly from PSE&G billing data — and benchmarked Fair Lawn against three comparable New Jersey towns' climate action plans. I led the team and took responsibility for organizing the work and presenting our findings.

The Problem

The Borough had never calculated its own emissions footprint, and its utility records weren't built for that kind of analysis — 93 separate PSE&G accounts, each billed in its own units (kWh for electricity, therms for natural gas), with no standardized format for comparing usage across properties or tracking it over time. Separately, Fair Lawn had no climate action plan of its own to guide next steps.

My Role

I led the team and was responsible for how the work got organized and delivered — coordinating who worked on which piece (data collection, the emissions methodology, the climate action plan research) and putting it all together into the final presentation to the committee and, later, the Mayor and Council.

Process

We used the GHG Protocol's emission-factor approach — GHG emissions = Activity data × Emission factor — applying it to two years of electricity and natural-gas billing data across all 93 accounts to convert raw usage into CO2-equivalent emissions. In parallel, we researched and benchmarked the climate action plans of three comparable NJ towns (Cherry Hill, Glen Rock, and Ridgewood) to see what a workable plan for Fair Lawn could look like.

From our presentation to the committee: raw PSE&G electricity/gas billing data (top) run through the emission-factor calculation to produce CO2e totals (bottom).

What I Built

• A full Scope 2 emissions calculation across 93 municipal accounts, 2018–2021, using the GHG Protocol methodology
• A comparative review of climate action plans from three benchmark NJ towns
• A set of recommendations for the Borough: energy-efficiency upgrades, stormwater management (rain gardens), an incentive-based recycling program, more tree planting, and support for sustainable transportation and community solar
• A parallel set of asks for other Fair Lawn stakeholders: the Board of Education and Chamber of Commerce measuring their own carbon footprints, and local businesses evaluating ESG factors

Our finding: CO2e emissions from electricity ticked back up in 2020–2021 after a dip, and natural-gas emissions rose steadily from 2019–2020 — the basis for our recommendations.

Outcome

We presented the findings and recommendations directly to the Mayor and Borough Council — the first time this kind of standardized, multi-year emissions analysis had been put in front of them.

Overview

Through TKS's Innovate Program, I built an AI model from scratch to detect deepfakes — media manipulated using machine learning — and used the project to go deep on adversarial ML, the same category of technique used to both generate and catch this kind of manipulation.

The Problem

Deepfakes are getting harder to catch by eye, and the techniques used to generate them and the techniques used to detect them are locked in the same arms race — as generation improves, detection has to keep up. I wanted to understand that relationship from the inside rather than just reading about it.

My Role

This was an independent build. I researched the underlying adversarial ML techniques myself and built the detection model from scratch, rather than starting from an existing library or pretrained detector.

Process

The model takes in media samples, extracts features associated with manipulation artifacts, and classifies each sample as authentic or manipulated. Researching how generation techniques work was necessary to understand what a detector should actually be looking for.

What I Built

• A trained deepfake-detection model
• Independent research into the adversarial machine-learning techniques used in both generating and detecting manipulated media

Recognition

Separately, as part of a 4-person team in the same TKS program, I worked on the Google x TKS — 2023 Fall Sprint Challenge, recommending "Google See": an internal platform concept to close product-knowledge gaps between Google's sales and engineering teams. Two slides from the recommendation deck we submitted:

From our submitted recommendation deck, with Muhammad Hunain, Ibrahim Rogers, and Natan Kramskiy.

Outcome

Ended the program with a working deepfake-detection model and a much deeper understanding of adversarial ML than I had going in — the kind that comes from actually building the thing, not just reading about it.