I'm a recent Master's graduate in Computer Science at the University of North Texas, and most of my work is about closing the gap between what AI can do in research and what it actually does in production. I've built autonomous agents, real-time tracking systems, medical imaging models, and platforms that let people talk to databases in plain English — and I care a lot about making all of it reliable, observable, and deployable.
Before grad school I spent 15 months as an Applied Scientist intern at CDK Global, working on ML systems for enterprise automotive retail. That's where I learned what it really takes to ship ML. I've also published research — a facial recognition paper at a Springer international conference — and I'm currently working on ASD detection from fMRI brain imaging.
✦ Peer-Reviewed · Springer Publication
Efficient Person Identification using Artificial Neural Networks
BVRITHCON-2023 International Conference · Published by Springer
My undergraduate thesis turned into a published paper. It compares Haar Cascade and MTCNN approaches for real-time face detection, evaluates LBPH recognition accuracy in a criminal identification context, and demonstrates the system working on live video. The research was accepted at an international conference and published by Springer.
Spent 15 months working on applied ML for one of the largest automotive retail platforms in the US. Most of my time went to model development, data preprocessing, and working with engineers to turn experimental work into something deployable. Learned what it actually takes to make ML reliable in production — not just accurate in a notebook.
Sep 2022 – Dec 2022
Business Model Creation — Open Innovation Program
GIET + University of California, Berkeley
Took part in UC Berkeley's Open Innovation Program during my final year of undergrad. Built a startup concept called MATSYA — a business model aimed at improving supply chain access for fishing communities. More about product thinking and stakeholder research than engineering, but genuinely useful for understanding how technology fits into real-world systems.
2024 – Present
Master of Science — Computer Science
University of North Texas, Texas USA
Currently focused on two things: the CNN-based ASD detection research project using 3D fMRI analysis with PyTorch, and building the kind of end-to-end AI systems that make up most of this portfolio. GPA 3.8.
GIET (Autonomous) · JNTU Kakinada, Andhra Pradesh, India
Bachelor's degree in Computer Science with specialization in AI & ML. Graduated with 8.0/10 CGPA. Published research at BVRITHCON-2023 International Conference during undergraduate studies.
Handles in seconds what used to take an engineer 20 minutes per issue
Got tired of watching GitHub issues pile up with no owner and no label. Built a LangGraph state machine that reads every new issue, classifies it, assigns it to the right person, and auto-closes anything that's been sitting open for 5 business days with no activity. The same issue always routes to the same person — deterministic assignment so nobody gets ping-ponged.
💼 Business Impact
Engineering teams typically spend 5–10 hours a week on issue triage and coordination. This agent reclaims that time entirely. For a team handling 200+ issues a month, faster assignment and automatic SLA enforcement means fewer project delays, fewer things falling through the cracks, and better engineering throughput — without adding headcount or process overhead.
Continuous production LLM monitoring on multi-cloud GCP + AWS
Production LLMs fail quietly — the model starts giving worse answers and nobody notices until users complain. This pipeline polls trace data from Arize every minute, runs each response through a Vertex AI judge model, and flags anything below threshold. Runs on GCP Cloud Run, deployed via Terraform, triggered on a schedule. Multi-cloud: built on AWS SageMaker, runs on GCP.
💼 Business Impact
Silent model degradation is one of the hardest problems in production AI. An LLM can start returning worse answers due to data drift or upstream changes — and teams don't find out until users complain or churn. Catching quality drops in minutes instead of days directly protects user experience, reduces incident response time, and helps AI product teams maintain the SLAs their businesses depend on.
Vertex AIAWS SageMakerDockerTerraformArizeCloud Run
Answers in 30 seconds what used to require a Jira ticket and a day's wait
Built for the situation where someone needs a quick data answer but the data team is busy. Type a plain English question, get back a SQL query and its results. The key piece is schema-grounded RAG — the model sees live table definitions, foreign keys, and business rules before generating anything. Runs entirely locally with Ollama and Gemma, so no data leaves the machine.
💼 Business Impact
The bottleneck in most data teams isn't analysis capacity — it's access. A product manager or operations lead who needs a quick data answer currently has to file a request, wait, and follow up. Self-service natural language querying removes that friction entirely. Faster access to data means faster decisions across sales, marketing, and operations — without growing the data team or training anyone in SQL.
Efficient Person Identification via Neural Networks
Published at BVRITHCON-2023 — Springer International Conference
Started as my undergraduate thesis and turned into a published paper. Detects faces through two pipelines — Haar Cascade for speed, MTCNN for accuracy on difficult angles — and matches them against a criminal database with real-time alerts. Green for no record, red for a criminal history, orange for unknown. The paper compares both detection approaches on live video footage.
💼 Business Impact
Manual identity verification doesn't scale. A security guard checking IDs at a busy entrance processes a few hundred people per hour at best. An automated system handles thousands, continuously, without fatigue. For enterprises, airports, or logistics hubs, this means lower staffing costs, faster throughput, and a timestamped audit trail of every identity event — something manual checks can never provide. The criminal record lookup layer converts a passive access control tool into an active security asset.
Turns supplier receipts into live inventory updates — no manual entry
Vending operators spend more time on admin than they should. Upload a receipt photo or PDF and the system reads it with OCR, updates warehouse stock automatically, and flags machines that need restocking before they run out. Revenue is calculated from actual receipt costs — not catalog estimates — so the profit numbers reflect reality. Built with React, FastAPI, PostgreSQL, and Supabase.
💼 Business Impact
Vending operators typically lose 15–25% of potential revenue to stockouts and product expiry. A demand-driven restock schedule directly cuts both. The OCR receipt pipeline eliminates 5–10 hours of manual data entry per week for a mid-size operator managing 50+ machines. Having profit calculated from actual invoice costs — not catalog estimates — lets operators see which machines and products genuinely earn their place, giving them the margin visibility to make real business decisions.
Early detection from brain scans — no manual feature engineering needed
Ongoing research using 3D CNNs on resting-state fMRI data. The challenge is that fMRI is volumetric — 3D spatial data across time — so we treat temporal frames as input channels. Evaluation is deliberately focused on Recall, not just Accuracy: in medical screening, a missed case costs far more than a false positive. Subject-level data splits prevent leakage across train and test sets.
💼 Business Impact
ASD affects 1 in 36 children in the US, but access to specialist diagnosis is unevenly distributed. A scalable neuroimaging-based screening tool could extend diagnostic reach to underserved populations, reduce wait times in overburdened clinics, and flag high-probability cases for priority review. The research-grade evaluation framework being built here is designed to meet the bar clinical deployment would require.
One operator monitoring 10+ live camera feeds simultaneously
Significantly upgraded an open-source multi-camera tracking system — replaced Darknet and TensorFlow 1.14 with YOLOv8, added vehicle intelligence (color detection, plate OCR, vehicle type classification), and fixed a concurrency bug where shared tracker state was corrupting IDs across camera threads. Every camera gets its own Deep SORT instance now. Settings update live without restarting the server.
💼 Business Impact
Traditional CCTV monitoring requires one operator per camera at meaningful scale — expensive, and still prone to human fatigue errors. One operator using this system can cover 10+ feeds simultaneously with automated event detection handling the rest. For retail, smart cities, or traffic management, the vehicle intelligence layer goes further: structured data on vehicle counts, types, and flow times feeds directly into operations dashboards, turning passive infrastructure into a genuine business analytics asset.
Professional-level Microsoft certification validating expertise in designing and implementing AI solutions — natural language processing, computer vision, and conversational AI on Azure Cognitive Services.
Covered core ML concepts and AWS services for building, training, and deploying machine learning models at scale — including SageMaker and the AWS AI/ML stack.
Python applied to data science workflows — data manipulation with Pandas and NumPy, visualisation, and the analytical foundations that underpin every ML project.
Foundation-level AWS certification covering core cloud concepts, global infrastructure, billing, security, and the key services that power modern cloud-native applications.
Microsoft Technology Associate certification in Python — covering core programming logic, data structures, OOP, and scripting fundamentals that form the backbone of all AI/ML engineering work.
Grew up playing cricket. Love the strategy as much as the sport itself — still play whenever I get the chance.
🏸
Badminton
Good for staying quick and focused. Fast-paced enough to actually be a workout, not just exercise.
💪
Fitness
Consistent gym routine. Helps me stay sharp, especially during long project sprints.
🎬
Movies
Mostly sci-fi and thrillers. Good storytelling teaches you something about explaining complex things simply.
🎨
Drawing
Sketch occasionally — mostly character art. A different kind of problem solving than writing code.
✈️
Travelling
Have crossed a few country borders so far. Every new place changes how you see things a bit.
Let's Build Together
Get In Touch
Open to Opportunities
Recent MS CS graduate at UNT. Available for AI/ML internships, research collaborations, and full-time roles. Let's create something extraordinary together.
