Naman Behl
In ninth grade, I ran a survey and stayed up all night looking for patterns in the responses. That curiosity became a habit: question the obvious, follow the data, and build something useful from what emerges. Today, at Illinois, I apply it across AI research, data engineering, and products designed for real teams.
Built different.
Thinks in data.
Born in Chandigarh. Grew up in Dubai. Now at the University of Illinois studying Information Sciences and Data Science, with minors in Computer Science and Economics.
Last summer I built an NLP-powered analytics agent at Kinesso that marketing teams used daily to query large datasets in plain English and get real-time strategy insights. Before that I was inside a Security Operations Center in Dubai watching enterprise systems operate under real pressure.
I also represented the UAE nationally in tennis, captained a cricket team, and published a poetry collection at seventeen. The through-line: I like understanding how things work and finding the pattern underneath.
Where I've worked.
What I can now build end to end.
Not a tool list—a set of production capabilities earned by shipping, debugging, and operating a real system.
Built Claude classifiers with schema-constrained outputs, deterministic fallbacks, cost-aware model routing, and uniform validation that rejects malformed or fabricated data before persistence.
Integrated Adjust, Meta, YouTube, Apify, and Google Sheets into a scheduled pipeline; evolved a Supabase/Postgres schema through real migrations and verified integrity from ingestion to rendered output.
Shipped a FastAPI service with OAuth2, session state, and REST chat; migrated scheduling from local cron to GitHub Actions; managed OAuth scopes, key rotation, and service-account credentials.
Built five production dashboards and a real-time chat UI, root-caused a silent precedence bug, reconciled divergent Git histories, and documented the system differently for engineers and stakeholders.
{{ job.title }}
{{ pt }}
When AI judges human conflict.
Professor Meicen Sun · UIUC School of Information Sciences
A broader study of how AI models interpret interpersonal conflict, using 120 real conversations from r/AmIOverreacting. GPT-4o judged whether Person B overreacted under control and modified-framing conditions, and predicted how conversations would continue from their first 60%.
My responsibility was the layer the model depended on: transcript quality. Because the model saw typed transcripts rather than the original screenshots, every missing emoji, altered word, or formatting inconsistency could change the emotional signal—and therefore the judgment.
Screenshot-to-transcript quality evaluation
Built an ML-assisted evaluation pipeline using GPT-4o vision to compare each conversation screenshot with its transcript across two dimensions:
Accuracy — words, punctuation, emojis, and message boundaries.
Consistency — whether the same transcription schema was applied across all posts.
Accuracy was strong. Consistency was the hidden risk.
Emoji handling was the largest systematic inconsistency: some transcripts retained emotional cues while others dropped them entirely. In a study measuring perceived overreaction, that missing tone is not cosmetic—it can affect the model's conclusion.
The audit also isolated seven corrupted transcript files caused by upstream transcription errors before they could contaminate downstream analysis.
Things I've built.
Let Data Speak.
A publication where I interrogate industry assumptions with data. No filler — just the numbers and what they actually say. {{ issueCount }} issues and counting. The archive syncs directly from Substack.
Work in progress.
Restaurant Location Intelligence Tool
A full-stack data pipeline and location intelligence tool that helps entrepreneurs identify the optimal neighborhood to open a restaurant. Given a cuisine type and budget, the system ranks neighborhoods by success likelihood across four dimensions: competition density, demographics, foot traffic potential, and budget fit.
The full picture.
{{ b.title }}
{{ b.desc }}