# Kunal Mahato

*Samsung · Turvo · Virginia Tech*

> Engineer who ships, now with AI in the loop.

- **Based:** San Francisco Bay Area · open to relocation
- **Availability:** Open to full-time roles (ET (GMT-4))
- **Status:** Open to work
- **Now:** Benchmarking frontier LLMs on sensitive data

### Pitched for

- *As AI Engineer:* AI Engineer building agentic RAG pipelines on Gemini, OpenAI, LLaMA, with 5+ yrs shipping production code at Samsung and Turvo before grad school.
- *As Forward Deployed Engineer:* Forward-deployed energy: ship fast, sit with customers, demo, iterate. 5+ yrs production code, now building AI products end-to-end.
- *As Software Engineer:* Software engineer who ships end-to-end: DB schema, API, UI, deploy. Android framework code on 1.48B Samsung devices, platform microservices at Turvo, now building AI systems at a VT research lab.

## 01 · Impact — By the numbers

- **1.48B+** — Devices reached at Samsung. 3 yrs on Android SystemUI: Quick Panel, Notifications, Status Bar
- **216K+** — Records migrated. Zero data loss · MySQL → MongoDB · Turvo Intermodal launch
- **99.9%** — Uptime. UPass Manager · 10K+ student records on Kubernetes
- **50%** — API speedup. Optimized GraphQL/REST endpoints on UPass Manager
- **80%** — F1 classification. Agentic RAG pipeline at VT research lab
- **3.93** — GPA at Virginia Tech. M.Eng. Computer Science, May 2025

## 02 · Why you should hire me

### As AI Engineer

I'm not an AI person who learned to code. I'm an engineer with 5+ yrs in production (Samsung SystemUI, Turvo platform) who pivoted into AI when it became clear it would reshape how software gets built. That means I ship LLM features that actually work in production, with evals, guardrails, and the plumbing around the model, not just the prompt.

**Stack I reach for:** OpenAI / Gemini / LLaMA · PyTorch + SentenceTransformers · Agentic RAG · LLM eval harnesses · FastAPI + Next.js

**Proof:**
- Agentic RAG benchmarking Gemini + OpenAI + LLaMA on FRIES consent classification
- DoDEx: hackathon RAG for Defense Contract Notices

### As Forward Deployed Engineer

This is the role I'm optimizing my career for. I want to sit close to customers, demo what I build, iterate on feedback, and be the front face of the product. I've already done a version of it: at Turvo I owned customer compliance for 10+ enterprise clients; at Samsung I flew to HQ in Seoul to ship a feature for a flagship launch. I learn domains fast (supply chain, Android internals, defense contracting, gentrification modeling) and I ship demos under 36-hour hackathon pressure.

**Stack I reach for:** Python + TypeScript for speed · React / Next.js for demos · FastAPI for backend · LLM tooling for AI-assisted workflows

**Proof:**
- Intermodal launch: 216K records migrated, zero data loss
- Shipping the world's first foldable at Samsung HQ
- DoDEx: hackathon RAG for Defense Contract Notices

### As Software Engineer

End-to-end ownership plus production rigor. I've built across the stack: Android framework (Java), backend platform services (Python/Node, Elasticsearch, MongoDB, MySQL), and modern frontend (React/Next.js/Redux). At Samsung I optimized SystemUI code running on 1.48B devices (memory down 20%, crash rate down 85%). At Turvo I architected ETL that migrated 216K records with zero data loss. I write tests (95% coverage on Review Me), profile before optimizing, and understand that the boring parts (observability, CI/CD, rollback plans) are what make features actually ship.

**Stack I reach for:** Java / Python / TypeScript · React / Next.js · Node / FastAPI / Spring · MySQL / MongoDB / Elasticsearch · K8s + AWS + CI/CD

**Proof:**
- Cut battery drain 20% on 1M+ low-end Samsung devices
- Intermodal launch: 216K records migrated, zero data loss
- UPass Manager: 99.9% uptime for 10K+ students

## 03 · Experience — Where I've shipped

### AI Research Engineer at Virginia Tech

*Remote · Jul 2025 – Apr 2026*

**Stack:** AI · RAG · Python · PyTorch · LLMs

- Built data-processing pipelines for LLMs in Python integrating OpenAI, Llama, and Gemini APIs through implementation of data-selection strategies including seed-based, embedding-based, alongside domain expert in the loop.
- Designed an agentic RAG pipeline (PyTorch + SentenceTransformers) with tool-use for top-K selection, improving relevance and recall and achieving ~80% F1 in classification tasks.
- Automated reporting with Pandas/NumPy, producing results across 12 model setups and reducing run time by 3x.

### Software Engineer at Turvo Inc.

*Hyderabad, India · May 2021 – Jun 2023*

**Stack:** Full-stack · Python · Node · Elasticsearch · ETL · CI/CD

- Enhanced global search and filter efficiency by 10% through advanced Elasticsearch indexing and GraphQL API implementation, improving data retrieval speed and accuracy for supply-chain operations.
- Architected a resilient Python-based ETL pipeline to launch Intermodal shipment mode, migrating 216,000+ multi-leg records between heterogeneous data stores (MySQL to MongoDB) with zero data loss during a critical schema redesign.
- Managed reorganization of the Rules Framework into a dynamic Business Rule Management System, enforcing complex role-mode-status compliance constraints across 10+ enterprise clients while decoupling business logic from the core codebase.
- Built Jenkins CI/CD pipelines for isolated dev setups, cutting deployment time by 40% and improving release reliability.
- Implemented a resilient Python library for HashiCorp Vault, ensuring secure secret injection with 99.99% uptime for 10 clients.

### R&D Engineer at Samsung HQ

*Suwon, South Korea · Dec 2019 – Dec 2019*

**Stack:** Android · Java · Framework

- Integrated backend UI components in Java using Android Framework with system services for foldable-specific notifications, contributing to the global launch of the world's first foldable phone at Samsung HQ.

### R&D Engineer at Samsung

*Noida, India · Jun 2018 – Apr 2021*

**Stack:** Android · Java · SystemUI · Performance

- Maintained and optimized full-stack codebase for 5 OS iterations, enhancing performance for 1.48+ billion devices.
- Optimized System UI (Quick Panel & Notifications) in Java for low end devices, cutting memory use 20% and battery drain 20%, reducing crash rate 85% and boosting user satisfaction to 4.4/5 across 1M+ devices.
- Handled 5 Application and Framework modules - Settings, Calendar, QuickPanel, Status-Bar and warning notifications.
- Upgraded the customer satisfaction metrics by 10% through implementation of Search feature for Quick Panel.
- Collaborated cross-functionally with teams to build backend data-porting solutions integrating 20+ network operators.

## 04 · Education — Academic record

### Virginia Tech — Master of Engineering

*Alexandria, VA, US · Aug 2023 – May 2025*

- **Major:** Computer Science & Applications
- **GPA:** 3.93 / 4
- **Coursework:** Software Engineering · Machine Learning · AI tools for SW Delivery · Cloud Computing · Usability Engineering · Intro to AI · Web App Development · Urban Computing

### NIT Allahabad — Bachelor of Technology

*Prayagraj, UP, India · Jul 2014 – May 2018*

- **Branch:** Computer Science & Engineering
- **GPA:** 3.53 / 4
- **Coursework:** Data Structures · Algorithms · DBMS · Distributed Systems · Operating Systems · Networks · Cryptography · OOM

## 05 · Key stories — Stories, for your next screen

Each story uses Situation · Task · Action · Result. The Ask about Kunal chat can expand any of these on the fly.

### Shipping the world's first foldable at Samsung HQ

**Situation:** In 2019, Samsung was racing to ship the world's first foldable phone. I'd been on Android SystemUI in India for a year.

**Task:** Ship foldable-specific notifications, a brand new interaction surface, in time for global launch out of Samsung HQ in Seoul.

**Action:** Moved to Seoul HQ for 3 weeks. Integrated backend UI components in Java with Android Framework system services, handling the new folded/unfolded states across the notification pipeline. Pair-coded with Korean teammates in a time-pressured environment.

**Result:** Shipped in time for the Galaxy Fold global launch. Code ran on every foldable shipped that year and is the foundation of the modern foldable notification stack.

### Intermodal launch: 216K records migrated, zero data loss

**Situation:** Turvo needed to launch Intermodal shipping (multi-leg, multi-modal freight: train, truck, ship). Schema didn't support multi-leg.

**Task:** Redesign the shipment schema AND migrate 216K+ existing live shipment records MySQL → MongoDB with zero data loss during a critical customer-facing window.

**Action:** Architected a resilient Python ETL pipeline with validation checkpoints, idempotent writes, and rollback capability. Ran migrations in batches with shadow reads to verify parity. Coordinated with customer success on migration windows for 10+ enterprise clients.

**Result:** Migrated 216,000+ multi-leg records with zero data loss. Intermodal shipping launched on schedule and became a core revenue product.

### Cut battery drain 20% on 1M+ low-end Samsung devices

**Situation:** Samsung's low-end Android lineup was getting poor reviews; users complained of lag, crashes, and battery drain in SystemUI.

**Task:** Profile Quick Panel + Notifications, find hotspots, ship optimizations without breaking 100+ device variants.

**Action:** Profiled with systrace, Perfetto. Identified wasteful redraws, synchronous I/O on the UI thread, and leaked listeners. Refactored draw paths, moved work to bg threads, added careful lifecycle cleanup.

**Result:** Memory −20%, battery drain −20%, crash rate −85%. User satisfaction rose to 4.4/5 across 1M+ devices. Pattern was later adopted across other modules.

### UPass Manager: 99.9% uptime for 10K+ students

**Situation:** VT capstone: the existing UPass (student transit pass) process was entirely manual, with 30+ min wait times, errors, and lost records.

**Task:** Build a Kubernetes-managed full-stack app that handles 10K+ student records, pickup notifications, NFC validation, and admin dashboards.

**Action:** Architected Next.js + React + Node.js on K8s. AWS Lambda automated S3 → RDS ingestion, cutting manual entry 95%. SES for pickup notifications. Optimized GraphQL/REST with JWT for role-based access, 50% faster. Socket.IO for 100ms NFC validation.

**Result:** 99.9% uptime. Student wait times halved. Manual entry down 95%. Shipped and in use for the Spring 2025 cohort.

### Agentic RAG benchmarking Gemini + OpenAI + LLaMA on FRIES consent classification

**Situation:** Research at VT extending prior Reddit-narrative work — measuring whether Gemini, OpenAI, and LLaMA can identify *which* of the six FRIES consent criteria is being violated in human-annotated real-world social-media stories.

**Task:** Design and lead the eval pipeline end-to-end: curate per-criterion benchmark with a team of Ph.D. psychologists in the loop; integrate Gemini + OpenAI + LLaMA under zero-shot and few-shot regimes; measure 3 K-shot example-selection strategies (random / similarity / handpicked); drive findings into manuscript.

**Action:** Agentic RAG pipeline (PyTorch + SentenceTransformers) with 768-dim sentence embeddings + cosine retrieval over annotated examples. Multi-provider LLM client (Gemini + OpenAI + LLaMA). Pandas/NumPy reporting across many model × criterion configurations.

**Result:** 80.4% F1 on 'any violation' identification (K=3 handpicked). Honest negative on per-criterion: majority of best-case cells below 50% F1, quantifying the gap that motivates the paper's call for human-expert intervention. Conference submission in progress.

### DoDEx: hackathon RAG for Defense Contract Notices

**Situation:** Bitcamp 2024, 36 hours. Government contracting data is unstructured, buried in PDFs, parsed today with brittle regex.

**Task:** Ship a working demo that extracts structured compliance data from Defense Contract Notices, with some measure of accuracy.

**Action:** Built an AI-driven ingestion pipeline in Python + RAG over Elasticsearch. Scraped a corpus of sample notices. Built a ground-truth validation framework to benchmark LLM outputs and prevent hallucinations.

**Result:** 30% higher extraction accuracy than regex baseline. Demoed live. Validation harness became the core idea I'm applying in my VT research now.

### Rules Framework → Business Rule Management System

**Situation:** Turvo's business rules (who can do what, when, in which shipment mode) were scattered across the core codebase. 10+ enterprise clients, each with custom compliance needs.

**Task:** Pull rules out of core code into a dynamic Business Rule Management System that could be configured per client without redeploys.

**Action:** Designed a rule schema covering role × mode × status. Built an evaluator with proper caching. Migrated existing hard-coded checks. Wrote tooling for the ops team to author and test rules.

**Result:** 10+ clients onboarded with custom rule sets. Core codebase decoupled from business logic, significantly reducing deploy risk for rule changes.

## 06 · Selected work — Projects

### UPass Manager

*Virginia Tech · Jan 2025 – May 2025 · Full-stack*

Kubernetes-managed Next.js web app for 10K+ student records with 99.9% uptime.

**Stack:** Kubernetes · AWS · React · Node.js · MySQL · NFC · Chart.js

- Architected Kubernetes-managed Next.js, React, Node.js web app ensuring 99.9% uptime for 10K+ student records.
- Cut manual entry 95% with AWS Lambda automating S3 → RDS ingestion, boosted pickups via SES notifications.
- Improved API speed 50% via optimized GraphQL/REST endpoints (Express, Apollo, JWT).
- Halved student wait times with a 100ms NFC validation pipeline using Socket.IO.

**Links:** [GitHub](https://github.com/mahatokunal/uPassManager)

### LifexAI

*University of Maryland · Bitcamp · Apr 2025 · AI*

AI detective web app: image understanding + transactional data for anomaly detection.

**Stack:** Gemini Vision · Capital One Nessie · Vercel · MongoDB · Tailwind

- Integrated Google Gemini Vision API, Capital One Nessie API, and real-time Chart.js dashboards, boosting anomaly detection accuracy 30%.
- Accelerated data loading 50% with React Query caching + Next.js APIs on Cloudflare Workers.
- 99.9% uptime via Vercel + automated CI/CD with Docker.

**Links:** [GitHub](https://github.com/nidhikamath2102/lifexai)

### Gentrification Prediction

*Virginia Tech · Aug 2024 – Dec 2024 · ML / Data*

ML pipeline fusing Census, Zillow, Eviction Lab data to predict gentrification risk.

**Stack:** Feature Engineering · Geospatial · ML · SHAP · SARIMAX

- Fused US Census, Zillow ZHVI, Eviction Lab data into unified time-series geospatial DB via ETL.
- Built binary classifiers (LR, RF, LightGBM) with 5-fold CV + GridSearchCV, achieving 0.77 ROC-AUC.
- Used SHAP to interpret PCA-reduced features, mapping back to socioeconomic variables.
- Built SARIMAX models to forecast 5-year SEIFA scores, flagging at-risk counties.

**Links:** [GitHub](https://github.com/The-Swapster/gentrification_classification_prediction)

### DoDEx

*Bitcamp Hackathon · Apr 2024 · AI*

AI ingestion pipeline for Defense Contract Notices, 30% more accurate than regex.

**Stack:** LLM · RAG · Elasticsearch · Web Scraping

- Python + RAG pipeline parsing unstructured Defense Contract Notices, 30% higher accuracy than regex.
- Built a ground-truth validation framework to benchmark LLMs and prevent hallucinations in mission-critical reporting.

**Links:** [GitHub](https://github.com/nidhikamath2102/DoDEx) · [Demo](https://devpost.com/software/ai-powered-extractor-for-defense-contract)

### Book Wonders

*Virginia Tech · Jan 2024 – May 2024 · Full-stack*

E-commerce platform: 1000+ transactions, JWT auth, 15+ REST APIs.

**Stack:** React · Redux · Java · MySQL · AWS

- Scalable e-commerce app with login, browsing, cart, and checkout, supporting 1000+ transactions.
- Responsive pages with React + Context API + Redux, improving load and data flow up to 30%.
- 15+ RESTful APIs in Java with MySQL integration.
- Deployed on AWS with JWT auth for secure sessions.

### Review Me

*Virginia Tech · Aug 2023 – Dec 2023 · Full-stack*

Peer review platform on MERN: hashtag search, upvotes, 95% test coverage.

**Stack:** MERN · JWT · Mocha · MVC

- Full-stack peer review platform on MERN with MVC architecture.
- Hashtag search, notifications, upvote/downvote, comments for content discovery.
- 95% test coverage with Mocha + Chai: unit and acceptance.
- Improved UX 20% with post editing, search, notifications, guest functionality.
- Secure auth via JWT with encrypted stateless communication.

**Links:** [GitHub](https://github.com/shekharmnnit/Runtime_Terror)

## Let's talk — Contact

*End of portfolio · begin conversation*

The fastest way to a reply is email. If you just want to browse, grab the PDF.

- **Email:** mahatokunal41@gmail.com
- **Resume:** https://www.kunalmahato.tech/KunalMahato_Resume.pdf
- **LinkedIn:** https://linkedin.com/in/kunalmahato
- **GitHub:** https://github.com/mahatokunal

---

*Kunal Mahato · Portfolio · 2026*