Aayush Maharjan — Machine Learning Engineer

00 · About

Research brain, production hands.

I'm a machine learning engineer at Smart Data Solutions, a US healthcare-tech company, where I own the document classification models behind the QuickClaim portal. I work the entire pipeline — scoping problems with stakeholders, designing and training models, then deploying them through CI/CD into live Java and Python services. Before ML, I spent three years as a software engineer, which is why my models don't just score well in notebooks — they survive production.

On the side: MSc research evaluating LLM and RAG systems on multi-hop QA, and internal training sessions on CNNs, Transformers and generative AI for the wider engineering team.

5+years building software

98%+classifier accuracy in production

2+years of ML shipped to prod

7Bparams squeezed onto one T4

01 · Experience

Where I've shipped.

Machine Learning Engineer

Smart Data Solutions MAY 2024 — PRESENT

Eagan, MN, USA · remote from Kathmandu

Own the document classification models for healthcare claims and attachments — EOBs, EOPs, UB-04s, CMS-1500s, ERAs, itemized statements, workers' comp — running live in the QuickClaim portal at 98%+ accuracy.
Replaced legacy rule-based routing with image classification pipelines built on CNNs, U-Net and Transformer encoders; built the OCR + NLP extraction that turns messy real-world scans into structured claim data.
Run the deployment side: Jenkins and GitLab CI/CD for model releases, an internal workflow for packaging reusable Python environments, and Python inference inside Java services via Jep — including CUDA memory management in the JVM.
Teach internal sessions on CNNs, NLP, Transformers and generative AI for the engineering org.

PyTorchCNN · U-NetTransformersOCR / NLPJenkins · GitLab CIJep · CUDA

Software Engineer

Smart Data Solutions JUL 2021 — APR 2024

Eagan, MN, USA · remote from Kathmandu

Built and maintained Java services for healthcare claim processing, working with onshore PMs to turn vague client requirements into shipped features.
Worked on Coupon Cabin, a high-traffic browser extension — mostly reliability and performance work.
Ran code reviews across the team and mentored newer engineers on debugging and design patterns.

JavaDistributed servicesPerformanceMentoring

Data Engineer Intern

Fusemachines MAY — JUL 2021

Kathmandu, Nepal

Pipeline design and deployment alongside senior data engineers using Hadoop, Spark and Airflow.

ML Engineer Intern

Treeleaf.ai NOV 2020 — FEB 2021

Lalitpur, Nepal

Classical ML through deep learning and computer vision; implemented object detection, face detection and face recognition models.

02 · Playground

Train a neural network. Right here.

A real multilayer perceptron — forward pass, backprop, gradient descent — written from scratch in vanilla JavaScript. No libraries, no WebGL, no tricks. Paint two classes of points, press train, and watch it carve out a decision boundary. Hover the plane to probe the network: the diagram lights up with live activations at every neuron.

class A · click class B · shift or right-click edge width = |weight|

Decision surfacetraining

the plane is empty —
click to paint points, or load a preset below

Network · weights & activations2 · 6 · 6 · 1

Optimizerbce loss

0steps

—loss

—accuracy

learning rate 0.08

training speed 24×

neurons / layer 6

hidden layers 2

Convolution Labslide a kernel over an image you draw — conv → ReLU → pool, live → Attention Labwatch a transformer decide which words matter — real Q·K^T softmax → CNN Guideconvolution worked by hand — LeNet & VGG counted parameter by parameter → LLM Guidethe whole story, tokens → RLHF — every matrix size, revision-ready →

03 · Projects & Research

Selected work.

Production

QuickClaim Document AI

The classifier suite that routes healthcare claims and attachments in production — CNN, U-Net and Transformer-encoder pipelines that replaced a brittle rule-based system, plus OCR/NLP field extraction feeding downstream claim automation. 98%+ accuracy on live traffic.

PyTorchU-NetTransformersOCR

MSc Research

RAG on Multi-hop QA

Benchmarked four setups on HotpotQA — plain LLM, standard RAG, RAG + chain-of-thought, and DeepSeek-R1-Distill as generator. Wrote the full eval pipeline: Answer / Supporting-Fact / Joint EM & F1, reasoning-trace stripping, all on T4s with 4-bit NF4 quantization and FAISS retrieval.

DeepSeek-R1FAISSBitsAndBytesHotpotQA

In development

Lumo SFA

A pharma sales-force automation backend: Django + Django Ninja on PostgreSQL with JWT auth, role-based access control across six roles, and an adapter layer designed for plugging in LLM features later. Built with uv, pytest and ruff.

Django NinjaPostgreSQLJWT · RBACuv · pytest

04 · Skills

The toolbox.

ML / Deep Learning

PyTorchTensorFlow / KerasHugging Facescikit-learnCNNsU-NetTransformersFine-tuning4-bit quantization

LLMs & RAG

RAG pipelinesEmbeddings · FAISSChain-of-thoughtDeepSeek-R1Mistral · QwenLangChainLLM evaluation

MLOps & Deployment

JenkinsGitLab CI/CDDockerLinuxJep (Python ⇄ Java)CUDA / GPU memoryEnv packaging

Languages

PythonJavaJavaScriptSQLC

Data & Backend

Hadoop · SparkAirflowREST API designDjangoFlaskNode.jsPostgreSQLSQL / NoSQL

Tooling

GitBashuvpytest · ruffSeleniumn8nSSH / WinSCP

05 · Education & Certs

Still learning, on purpose.

MSc in IT — AI Specialization

ISLINGTON COLLEGE, KATHMANDU · EXPECTED 2027

Deep Learning, NLP, Computer Vision, Generative AI, applied RAG research.

Bachelor of Information Management

NATIONAL COLLEGE OF COMPUTER STUDIES · 2016 — 2020

Kathmandu, Nepal.

Certifications

Deep Learning Specialization — DeepLearning.AI (Coursera)
Machine Learning: Tree-Based Models
MongoDB: Data Modeling

06 · Contact

Let's build something that actually ships.

I'm most at home where research meets production — if that's the kind of problem you have, my inbox is open.

[email protected] LinkedIn ↗

☏ +977 9808084972 ⌖ Kathmandu, Nepal · UTC+5:45 résumé.pdf ↓