Naman
Mittal
Software Engineer
Welcome aboard. This flight travels backward through time · from where I am now, to where it all began.
BNY
I build across the entire stack · from raw market data in Snowflake to the screens a trader watches. Touch any station on the line.
Data
SNOWFLAKE · HADOOP · WAREHOUSEMigrated financial datasets from Hadoop to Snowflake with Java-based services · a large-scale data infrastructure initiative that improved accessibility and query performance for every downstream analytics and ML workflow.
Toolkit
Everything I reach for to get from idea to shipped · languages first, then the data stacks, clouds, and libraries built around them.
LANGUAGES
DATA ENGINEERING
FRAMEWORKS
ML & LIBRARIES
CLOUD & DEVOPS
TOOLS
Projects
Every project below is a working interface. Scroll it, tap it, flip it, search it.
Treadmill screens forget your workout in 30 seconds. GymReader makes one photo remember it forever: Apple's built-in Vision OCR parses the screen entirely on-device, you review every confidence-graded value, then it's saved to Apple Health. Scroll · the phone advances through the workflow.
- 01 TREADMILL SCREEN
- 02 CAPTURE
- 03 OCR ON-DEVICE
- 04 REVIEW & SAVE
- 05 APPLE HEALTH
The treadmill screen. It forgets everything in 30 seconds.
A stick-figure flipbook engine, rendered limb by limb from math. Type any story · an LLM storyboards it from 17+ actions or writes new animation code on the spot · then the canvas draws 80 frames on a tactile page turn. The real app is embedded below · pick a template or write your own.
Nyai
REACT · REAL-TIME · ENGLISH · HINDIAn interactive Indian law practice platform · students analyze real landmark cases and argue them in a live courtroom, with AI grading each response server-side.
Curious about the codebase, the courtroom engine, or the curriculum catalog? Contact me for more details.
Agent Eval
DEEPEVAL · RAG · LLM TESTINGThe same RAG agents I build at work need to be proven reliable · so I architected a testing framework that interrogates them before users ever do.
RAG-based test cases simulate multi-turn agent interactions for conversational business analysis, then automatically score reasoning quality, guideline adherence, factual accuracy, hallucination detection, and task completion · across every release.
HbA1c Prediction
PYTHON · SCIKIT-LEARN · REGRESSIONMy Masters research in UC Riverside's CS department · predict HbA1c test results from daily blood-sugar readings using regression. Glucose windows become the features and scikit-learn models do the prediction · the best model lands within ±0.5% HbA1c, and it's the first ML treatment of this problem.
No prior work predicts HbA1c with machine learning · the existing methods are formula- or kinetic-model-driven
Predicts from readings patients already take daily · removes a separate clinic appointment
Linear regression hits a 0.50 mean absolute error using area-under-the-curve features
Built on the JCHR public dataset of type-1 diabetic patients · 50-day blood-sugar windows, using either raw readings or the area-under-the-curve as input features. Compared Linear, Decision Tree, and Random Forest regression on MAE · MSE · RMSE · linear regression wins across the board.
loanDepot
A summer inside mortgage fintech · where I learned that shipped beats perfect.
Payoffs & text quotes
Customer-facing RESTful web service in Angular 11 and .NET Core for requesting loan payoffs and texting quotes to customers.
Live-agent support chat
Real-time customer support chat in Microsoft 365 Dynamics using JavaScript and real-time communication protocols.
Performance engineering
Monitored and tuned 20+ performance metrics in Dynatrace · latency down, responsiveness up across the web app.
Ship measurable work
Enterprise SDLC, code review, CI/CD · and that every feature is a number somebody watches. Mine went the right way.
UC Riverside
Two years of going deep: the theory behind the systems I now build for a living.
Community & Impact
I build software for a living. Outside of it, I’ve spent years building something harder to measure: a community, an audience, and a place where people can find answers I once had to figure out myself.
What started as sharing my own experience moving from India to the US grew into a 1,300+ member community for international students. A place for the questions that rarely make it into brochures: choosing a university, navigating visas, finding internships, getting your first job, and figuring out life in a new country.
It started with an open notebook. I shared what I was learning, getting wrong, and figuring out along the way.
Moving from India to California changed the subject. The posts became field notes from studying, adapting, and building a life thousands of miles from home.
From UC Riverside to BNY Mellon and beyond, the journey became less about documenting a destination and more about sharing what happens while getting there.
The audience became a community. The community became a network. 14K+ on LinkedIn. 1,300+ in Discord.
No playbook. Just consistency, curiosity, and building in public.
Undergrad
Before software became my career,
it was curiosity.
The first program
A calculator on a machine that wasn't mine. It crashed. I fixed it. I was hooked.
The lab hours
Every spare hour in the computer lab · debugging by instinct before I knew the word for it.
The decision
By second year I was building things nobody assigned. That's when it stopped being a subject.
3-2-1 GoCheck
The first cockpit · machine learning on real documents, before it was fashionable.
OCR system at 98%
A Python OCR pipeline on pytesseract and TensorFlow YOLOv3 that reads and verifies documents automatically.
Document automation
Angular 7 + TypeScript tooling that verified and updated documents on its own · the manual process ceased to exist.
Pipeline busywork cut
Wired Clubhouse to Git repositories to drive the pipeline, removing 90% of manual user-story updates.