Currently
CS sophomore @ UT Dallas
Based in
Allen, TX
Open to
Summer 2027 internships
Updated
Nov 2026

Projects and proof

01 / Internlog: case study

Internlog

An anonymous internship-review platform with an explainable fraud and spam detector.

Node/Express Supabase Render

Free hosting: the first load can take a few seconds to wake up.

Built with Suhaas Voruganti.

What it is

Students browse and submit internship reviews, company ratings and interview experiences. Submissions are screened by a rule-based detector that scores five signals: lexical diversity, near-duplicate similarity, rating extremity, templated language and link spam. It shows moderators why a review was flagged instead of rejecting it automatically.

What I built

  • Moved storage from flat JSON files to Supabase Postgres with four tables and row-level security enabled.
  • Wrote the data-access layer so the frontend needed no changes after the migration.
  • Added moderation workflows, email notifications through the Brevo API, and rate limiting.
  • Did a full mobile pass: hamburger nav, responsive CSS, small-phone breakpoint at 380px.
  • Set up GitHub Actions CI/CD and CodeQL security scanning, plus uptime monitoring.
  • Ran a security hardening pass: admin-gated write endpoints, an XSS fix, a query fix that protects the database connection pool, 36 automated tests that run on every push, and every npm audit finding cleared (Nov 2026).
Demo data

A replica of the Internlog landing page with fictional demo companies: the tier pills and the search box really filter the cards, and the recent-reviews strip pauses on hover.

View data
Demo companies
CompanyTierIndustry
Meridian Cloud SystemsTier 3 · GoldTechnology
Cobalt AerospaceTier 3 · GoldAerospace
Lone Star AnalyticsTier 2 · SilverData
Brightline RoboticsTier 2 · SilverRobotics
Oakridge Community BankTier 1 · BronzeFinance
Pinecrest DigitalTier 1 · BronzeMarketing
Fig. 01 · Internlog home
Demo data

A fictional company page: the stats are computed from its four reviews (average 8.5/10), and sort and the role filter really reorder the cards.

View data
Meridian Cloud Systems (demo): average rating 8.5, difficulty 8.5
RoleRatingDifficultyPayPosted
Software Engineering Intern9/108/10$48/hr2026-09-02
Software Engineering Intern9/108/10$48/hr2026-08-30
Cloud Infrastructure Intern8/109/10$46/hr2026-08-24
Data Platform Intern8/109/10$44/hr2026-08-15
Fig. 02 · A company page with reviews
Demo data

The mobile landing page in a phone frame: open the menu, and drag the width handle between 320, 380 and 430 pixels to see the small-phone layout take over at 380.

View data
Viewport widthLayout
320pxSmall-phone layout
380pxSmall-phone breakpoint
430pxStandard phone layout
Fig. 03 · Internlog on a phone

Architecture

Node/Express 5 REST API; vanilla HTML/CSS/JS frontend with no framework; Postgres on Supabase via a pg connection pool over the IPv4-compatible session pooler (migrated from the original JSON-file storage); Render hosting with deploys on pushes to main and UptimeRobot pings to avoid free-tier spin-down; Brevo HTTPS email API instead of Gmail SMTP because Render's free tier blocks outbound SMTP ports; configuration through environment variables.

The browser runs a vanilla HTML, CSS and JavaScript frontend that calls an Express 5 REST API on Render. The API reaches Supabase Postgres through a pg pool on the IPv4 session pooler, and sends email through the Brevo HTTPS API. UptimeRobot pings the API, and GitHub Actions with CodeQL deploys on pushes to main. BROWSERvanilla HTML/CSS/JS EXPRESS 5 REST APIRender · env config PG POOLsession pooler (IPv4) SUPABASE POSTGRES4 tables · RLS on BREVO EMAIL APIHTTPS (SMTP blocked) UPTIMEROBOTpings: no spin-down GITHUB ACTIONSCI · CodeQL deploy on push to main direct connection is IPv6-only
Fig. 04 · Internlog architecture

Backend practices

Parameterized SQL; mapper functions from snake_case rows to a camelCase API; a shared asyncRoute wrapper for consistent error handling; idempotent INSERT … ON CONFLICT SQL for data fixes and restores; the company list built from one aggregated query instead of one query per company; clean URLs with a 301 from /page.html to /page.

Explainable fraud detector

A transparent, rule-based Stage 1 scorer (not a trained model) with no dependencies. It checks five signals: length and vocabulary diversity, near-duplicate similarity (Jaccard similarity on word 3-gram shingles against the company's existing reviews), rating extremity (a 1 or 10 paired with thin text), generic or templated phrasing, and links (with extra weight for commonly abused domains). It produces a 0–100 score, a level (low, medium or high) and plain-English reasons, stores all three with the review, and adds them to the moderation email. It never auto-rejects, so a human makes the final call. I chose a rule-based scorer over a separate ML microservice on purpose: every flag has to be explainable, and there's no labeled data yet to train on.

A submission passes validation and rate limiting, then the five-signal scorer, which checks length and vocabulary, near-duplicates, rating extremity, templated phrasing, and links and abused domains. The score, level and reasons are stored with the review and added to a moderation email with a ticket ID and a delete command. A human makes the decision; the system never auto-rejects. Scores under 30 are low, 30 to 59 medium, 60 and up high. SUBMISSION VALIDATION + RATE LIMIT FIVE-SIGNAL SCORER 1 LENGTH + VOCABULARY 2 NEAR-DUPLICATE (JACCARD) 3 RATING EXTREMITY 4 TEMPLATED PHRASING 5 LINKS + ABUSED DOMAINS SCORE, LEVEL +REASONS STORED MODERATION EMAIL(TICKET ID + DELETE COMMAND) HUMAN DECISION NEVER AUTO-REJECTS 03060100 LOWMEDHIGH 0-100 SPAM SCORE A submission passes validation and rate limiting, then the five-signal scorer, which checks length and vocabulary, near-duplicates, rating extremity, templated phrasing, and links and abused domains. The score, level and reasons are stored with the review and added to a moderation email with a ticket ID and a delete command. A human makes the decision; the system never auto-rejects. Scores under 30 are low, 30 to 59 medium, 60 and up high. SUBMISSION VALIDATION + RATE LIMIT FIVE-SIGNAL SCORER 1 LENGTH + VOCABULARY 2 NEAR-DUPLICATE (JACCARD) 3 RATING EXTREMITY 4 TEMPLATED PHRASING 5 LINKS + ABUSED DOMAINS SCORE, LEVEL +REASONS STORED MODERATION EMAIL(TICKET ID + DELETE COMMAND) HUMAN DECISION NEVER AUTO-REJECTS 03060100 LOWMEDHIGH 0-100 SPAM SCORE
Fig. 05 · Detector pipeline, five signals and score bands
Demo data

A fictional low-effort review and the moderation email it triggers. The score is computed live by the same scorer as the detector demo below; hover a reason to see which fields set it off.

View data
SignalPointsWhy
Length + vocabulary+30Extremely short review text (<20 chars)
Templated phrasing+15Generic/templated phrasing ("none")
Rating extremity+15Extreme rating (10/10) paired with thin review text
Near-duplicate0No overlap
Links0No links
Total60 / 100HIGH: flag only, a human decides
Fig. 06 · Flagged moderation email

Try the detector

A simplified demo, not the production code.
Score -- waiting
  • Type a review, or try one of the examples.

Security

Tiered rate limiting (general, write, admin) with trust proxy set correctly behind Render; server-side validation (rating 1–10, 250-word limit per field); an admin key, compared in constant time, required to create or edit companies and to delete reviews (401 without it); escapeHtml() on user-controlled output, including quotes, which closed an attribute-injection hole in the company dropdown (the first XSS fix came from a CodeQL finding); row-level security enabled on the tables with no policies and server-only access; no user accounts, so little personal data; Privacy Policy, Terms and SECURITY.md plus a required policy-agreement checkbox; CodeQL and Dependabot scanning; 36 automated tests on every push.

Moderation and product

An email for every new review and report; a report modal for community flagging; a scope guardrail that keeps the site internships-only; company verification and tiering (Bronze, Silver, Gold) with leaderboards; search, role filtering and sorting; loading spinners; SEO and Open Graph tags; launch README with badges.

How I work

Prototype in a sandbox, then hand off tested code. I ran the scorer against realistic good and junk reviews before it touched the repo, and verified each step with syntax checks, grep, a local boot and a live end-to-end test (a thin review got flagged, a real one did not). I pull before pushing and resolve merge conflicts with the security fix preserved. After an accidental mass deletion, I rebuilt the data from the moderation emails.

What I learned

Supabase's direct connection is IPv6-only, so Render's free tier needs the session pooler. Render's free tier blocks SMTP ports, so email has to go over an HTTPS API.

What I'm improving next

  • Add automated database backups.
  • Record moderator decisions. Deleting a review currently removes its row, so the outcome is lost. Keeping decisions next to the stored score and reasons is the prerequisite for a labeled dataset and a future trained stage.
  • Node.js
  • Express 5
  • Vanilla JS
  • Supabase Postgres
  • Render
  • GitHub Actions
  • CodeQL
02 / FlyRank ML internship track

Search Content Decline Analysis

A reproducible pipeline that ranks 30,000 pages for content refresh and tests whether a model beats a simple rule.

  • Built a Python, pandas and scikit-learn pipeline on an anonymized 30,000-page search dataset.
  • Compared a hand-written rule against decision trees using Precision@K. At K=50 the depth-2 and depth-3 trees scored 0.60 and 0.72 against 0.68 for the rule, measured in-sample.
  • Tested whether keyword search volume predicts impressions. The correlation was about 0.00 across all 30,000 pages and across the 21,994 pages with 100+ impressions.
  • Demonstrated how target leakage inflates results.
Precision@50, in-sample
ApproachPrecision@50
Hand-written rule0.68
Depth-2 tree0.60
Depth-3 tree0.72
Precision@50 · in-sample, so this doesn't show the trees beat the rule

Honesty note These numbers are in-sample, so they don't show that the trees beat the rule.

Python pandas NumPy scikit-learn Google Colab

Program: FlyRank self-paced ML internship track (Nov 2026 to present).

Repo on GitHub
03 / Open-source Claude skill

Checkpoint Study Path

A free, open-source Claude skill that turns your slides into a study plan and a mastery-gated quiz.

It builds a study plan from course materials, computes the score needed, groups questions into checkpoints instead of chapters, brings missed questions back three questions later, and checks answer keys with a script. MIT license, built for students.

Requires a paid Claude plan.

A 30-second pixel-art demo with a music soundtrack and no narration. Every message is shown as on-screen text. It opens with "Exam in 9 days" and a wall of lecture decks and homework, then the title "Checkpoint Study Path, an exam-prep skill for Claude". A study plan shows the score needed on the final and a week-by-week timeline. A checkpoint map shows a cleared checkpoint unlocking the next, and a missed question coming back three questions later. A mixed review draws ten random questions from cleared checkpoints, and a review loop follows. It ends on the slash command and the GitHub link.

Repo on GitHub
04 / Summer 2026 internship

MindSoft Solutions internship

IT and Web Development Intern at MindSoft Solutions LLC, June to August 2026 (Allen, TX, hybrid).

I designed and built the company's official website from scratch, deployed it through Git, and wrote client-facing technical content on cybersecurity, cloud and network infrastructure services. Details are under Experience below.

Repo on GitHub
05 / Competitions

Competitions

National Cyber League, Fall 2026 Individual Game

Ranked 420th of 5,959 (top ~7%). Scored 2,335 of 3,000 points at 84.9% accuracy (141 of 166 correct, 0 skipped) across nine categories in a solo weekend competition.

Results are official after NCL's audit announcement.

NCL Fall 2026 points by category
CategoryPointsOut of
Log Analysis300300
OSINT385390
Password Cracking275335
Cryptography290375
Forensics245300
Network Traffic Analysis230300
Exploitation and Enumeration210300
Scanning and Reconnaissance200300
Web Application Exploitation100300
Next focus: web application exploitation (100/300).

Survey: 100 / 100 (not a competition category)

From the Cyber Skyline report

My NCL Fall 2026 Individual Game report as a terminal: rank and totals first, then five charts you can inspect. My values are exact; other participants' values are approximate readings of the report.

View data
NCL Fall 2026 Individual Game
MeasureValue
Rank420th of 5,959 (top ~7%)
Points2,335 / 3,000
Accuracy84.9%
Completion86.5%
Correct141 / 166
Skipped0
Hints3
Completion: OSINT100%
Completion: Log Analysis100%
Completion: Network Traffic Analysis94%
Completion: Forensics94%
Completion: Cryptography93%
Completion: Exploitation and Enumeration90%
Completion: Password Cracking76%
Completion: Scanning and Reconnaissance67%
Completion: Web Application Exploitation27%
Fig. 07 · NCL Fall 2026 report

CSAW CTF 2026, team captain

Captained a 3-member UT Dallas Computer Security Group team ("hugh_hack") in a 48-hour jeopardy-style CTF. The team placed 260th, roughly the top 20 to 25% of scoring teams, solving challenges in six categories: Web, Forensics, Cryptography, Misc, OSINT and Reverse Engineering.

Team result

CSAW CTF 2026 as a status card: the team placed 260th, in the top 20–25% band of a field of 1,000+ scoring teams. The totals at the bottom are the team's.

View data
MeasureValue
RoleTeam Captain, hugh_hack (UTD Computer Security Group)
Place260th, top 20–25%
Field1,000+ scoring teams, historically
Categories solvedWEB, FORENSICS, CRYPTO, MISC, OSINT, REV
Team points912
Challenges solved (team)10
Duration48 hours, Sept 2026
Fig. 08 · CSAW CTF 2026
06 / Certifications and honors

Certifications and honors

ItemIssuerDate
CS50: Introduction to Artificial Intelligence with PythonHarvard UniversitySept 2026
Python Programming EssentialsRice UniversityJul 2026
AP Scholar with DistinctionCollege Board2024
National Society of Leadership and Success (inductee)NSLSMay 2026
CompTIA Security+: in progressCompTIATarget January 2027
07 / Experience, education, skills

Experience, education, skills

Journey

  1. 2024AP Scholar with Distinction
  2. Aug 2025University of Texas at Dallas, B.S. in Computer Science
  3. Jun–Aug 2026MindSoft Solutions LLC, IT and Web Development Intern
  4. Nov 2026FlyRank, Machine Learning Intern (self-paced track)

IT and Web Development Intern

MindSoft Solutions LLC · Allen, TX (hybrid) · June to August 2026

  • Designed and built the company's official website from scratch with HTML, CSS, JavaScript and GitHub Pages, turning IT and cybersecurity service offerings into a clear, professional site.
  • Deployed a responsive multi-page site managed end to end through Git.
  • Built a mobile navigation menu and a working contact form, and wrote client-facing technical content on cybersecurity, cloud and network infrastructure services.

Machine Learning Intern (self-paced track)

FlyRank · Nov 2026 to present

Search Content Decline Analysis project (see 02).

Leadership and activities

  • Team captain, CSAW CTF 2026, UT Dallas Computer Security Group (Sept 2026).
  • Member of ACM, the Artificial Intelligence Society (AIS) and the Computer Security Group (CSG) at UT Dallas.

Education

University of Texas at Dallas, Richardson, TX · B.S. in Computer Science with a Certificate in Cyber Defense (in progress) · Expected May 2029.

Relevant courses: Discrete Mathematics (CS 2305, current); Programming Fundamentals in C++ (CS 1436, Spring 2027).

Skills

GroupSkills
LanguagesPython, JavaScript, C++, MATLAB, HTML/CSS, SQL
Web and toolsNode.js, Express, PostgreSQL (Supabase), Git and GitHub, GitHub Actions (CI/CD), GitHub Pages
ML and datascikit-learn (decision trees), pandas, NumPy, model evaluation (Precision@K, train/test leakage), Google Colab
Security and AICTF skills (web, forensics, cryptography, OSINT), AI foundations (search, knowledge representation, neural networks), CompTIA Security+ (in progress)
Back to the top