AI/ML engineer building computer vision, reinforcement learning and the infrastructure that runs them.

Chitwan & Kathmandu, Nepal · Open to remote roles, AI/ML internships and research collaborations

01 Selected work

Computer vision · Android

KU Krishi

2024 · Kathmandu University · Published under bala.ai

Problem
Farmers in remote areas lose crops to diseases like late blight before an agronomist can ever see them.
What I built
An offline-first Flutter app that runs quantized YOLOv8 / MobileNet models on the phone and explains the result in Nepali.
Result
Semi-finalist at the Nepal ICT Award 2024, and a published app farmers can use with no signal.
  1. Leaf photo
  2. On-device YOLOv8
  3. Disease diagnosis
  4. Advice in Nepali
Fig. 02 — Inference runs on the phone, so a diagnosis never waits for the network.

Flutter · YOLOv8 · TensorFlow Lite · OpenCV · FastAPI · Firebase

Reinforcement learning · IoT · In progress

Hydroponics AI

2026 · STEAM research grant · DFRobot hardware sponsorship

Problem
Hydroponic farms dose nutrients on fixed thresholds, which wastes nutrients and caps yield.
What I built
Six sensing and control workflows on two ESP32 boards, a FastAPI service, and an RL policy (PyTorch, Gymnasium) that decides dosing, with live video over LiveKit.
Result
Funded by the Ministry of Education, Science and Technology’s STEAM grant. DFRobot sponsored the sensors after I pitched the project by email.
  1. Sensors
  2. 2× ESP32
  3. MQTT
  4. FastAPI
  5. RL policy
  6. Pumps & dosing
Fig. 03 — The control loop: sensor readings in, dosing decisions out.

PyTorch · Gymnasium · ESP32 (C++) · Raspberry Pi · MQTT · FastAPI · WebSockets · LiveKit · Flutter

More work

Mobile · GIS · bala.ai

TrekSathi

Trekking companion for Nepal’s mountain routes with no cellular coverage: offline trail maps, elevation profiles, severe-weather alerts and AI-assisted route suggestions.

  1. OpenWeatherMap
  2. Django REST · MongoDB
  3. Offline map cache
  4. Flutter app

Flutter · Google Maps SDK · Django REST · MongoDB · Firebase

Mobile · Marketplace · bala.ai

Like Tiffin

Meal subscriptions connecting students and office workers with home cooks, hotels and restaurants, with real-time orders and delivery-radius matching.

  1. Flutter client
  2. Firestore sync
  3. Node.js admin
  4. My bare-metal server

Flutter · Firestore · Node.js · Google Maps SDK

MLOps · 2025 · Kathmandu University

AIHPC GPU cluster

Multi-node Kubernetes cluster for deep-learning research in KU’s Department of AI, load-balanced with HAProxy and monitored with Prometheus and Grafana.

  1. Bare-metal nodes
  2. Kubernetes
  3. HAProxy
  4. RTX 3080 Ti GPUs

Lesson: consumer GeForce cards lack MIG partitioning, so stable multi-tenant GPU sharing wasn’t possible on this hardware.

Kubernetes · HAProxy · Docker · CUDA · Ubuntu Server

Infrastructure · 2022

Self-hosted production server

A public-facing bare-metal Linux server I built and ran to host the Like Tiffin backend instead of paying for a cloud provider.

  1. Internet
  2. Static IP · firewall
  3. HAProxy · TLS
  4. Docker services

Linux · HAProxy · Docker · MinIO · PostgreSQL · Node.js

Web · Sep 2023 – May 2024

PixShop

Multi-store e-commerce platform: one shared Next.js storefront and Node.js backend serving independent shops, each with its own catalogue, inventory and orders.

  1. Independent shops
  2. Next.js storefront
  3. Node.js services
  4. Per-shop data

Next.js · React · Node.js

02 About

Portrait of Safal Kumar Shrestha

I’m studying for a B.Tech in Artificial Intelligence at Kathmandu University (graduating 2027), and I build ML systems end to end — the model, the servers and GPU clusters it runs on, and the app people actually use.

I publish my apps under bala.ai and work on AI research and community projects with the KU Artificial Intelligence Club. When a project needs hardware, I ask for it: DFRobot sponsored the Hydroponics AI sensors after I pitched them by email.

Recognition

  • 2026

    STEAM Research Grant — Ministry of Education, Science and Technology, Nepal · Hydroponics AI

  • 2026

    DFRobot hardware sponsorship · Hydroponics AI

  • 2024

    Nepal ICT Award, semi-finalist · KU Krishi

  • 2024

    AI Crusade hackathon, 3rd place · KUAIC

03 Experience

Work

  1. 2025

    HPC Core Engineer, AIHPC initiative

    Department of Artificial Intelligence, Kathmandu University

    Led the build of the department’s shared GPU cluster, from bare metal to Kubernetes scheduling and HAProxy load balancing.

  2. Feb – Jul 2022

    Software Engineer, Android

    Shop2App Corp · United States, remote

    Built native Android (Java) and Flutter modules on Shopify’s GraphQL and REST APIs so merchants could ship their stores as branded Android apps. Improved checkout flows and automated Play Console releases.

  3. 2022 – now

    Member, KU Artificial Intelligence Club (KUAIC)

    AI research, workshops and community projects with peers at Kathmandu University.

Education

  1. 2022 – 2027

    B.Tech in Artificial Intelligence

    Kathmandu University, School of Engineering · Dhulikhel · expected 2027

  2. 2018 – 2022

    Diploma in Computer Engineering

    Nepal Polytechnic Institute · Bharatpur · 81.62%

04 Skills

Machine learning

  • PyTorch
  • TensorFlow / TFLite
  • YOLOv8 · OpenCV
  • scikit-learn
  • Gymnasium · PINNs
  • LangGraph · MCP

Infrastructure

  • Kubernetes
  • Docker
  • HAProxy
  • Linux administration
  • NVIDIA GPU profiling
  • CI/CD

Mobile

  • Flutter · Dart
  • Android (Java / Kotlin)
  • Firebase
  • Google Maps SDK
  • Play Console

Backend

  • FastAPI
  • Django REST
  • Node.js
  • MongoDB · MySQL · PostgreSQL
  • WebSockets · LiveKit

Hardware

  • ESP32
  • Raspberry Pi
  • Arduino
  • DFRobot sensors
  • MQTT

Languages

  • Python
  • SQL
  • Java · Kotlin
  • Dart
  • C / C++

05 Research interests

  1. Reinforcement learning for closed-loop control

    Q-learning and PPO policies that keep pH and EC on target against environmental drift in hydroponics.

  2. Deep RL for autonomous driving

    PPO and SAC agents for motion planning and obstacle avoidance in simulated environments.

  3. On-device computer vision

    Quantized MobileNet and YOLO models for low-latency, offline diagnosis on low-end phones.

  4. Physics-informed neural networks

    Embedding the differential equations of heat, fluid flow and diffusion into model training.

  5. LLM agents

    Stateful LangGraph workflows, local tool-calling models, memory and MCP, applied to trek planning.

06 Outside work

When I’m not writing code I’m usually on a trail. Nepal’s mountains are where I switch off — and the reason I built TrekSathi.

  • 4,130 m

    Annapurna Base Camp, North ABC route · 2025/26

  • 4,500 m

    Mardi Himal viewpoint, via High Camp · Kaski

Trekkers walking a moss-covered forest trail
Forest trail, Mardi Himal
Snow-covered ridge with lodges at the Mardi Himal viewpoint
View Point
High Camp direction sign surrounded by rhododendron leaves
High Camp
Wooden Forest Camp signpost among trees on the Mardi Himal trail
Forest Camp

07 FAQ

  1. Who is Safal Kumar Shrestha?

    Safal Kumar Shrestha is an AI/ML engineer from Nepal, studying for a B.Tech in Artificial Intelligence at Kathmandu University (expected 2027). Safal builds machine-learning systems end to end: computer vision and reinforcement learning models, the GPU clusters and servers they run on, and Android apps published under the label bala.ai.

  2. What are Safal Kumar Shrestha’s best-known projects?

    KU Krishi, an offline app that diagnoses crop diseases on the phone and advises farmers in Nepali (Nepal ICT Award 2024 semi-finalist); Hydroponics AI, a reinforcement-learning nutrient-dosing system funded by a 2026 STEAM research grant, with sensors sponsored by DFRobot; TrekSathi, an offline trekking companion; and the AIHPC GPU cluster at Kathmandu University.

  3. What technologies does Safal work with?

    Machine learning with PyTorch, TensorFlow Lite, YOLOv8 and OpenCV; infrastructure with Kubernetes, Docker and HAProxy; mobile apps with Flutter and Android (Java, Kotlin); backends with FastAPI, Django REST and Node.js; and hardware with ESP32, Raspberry Pi and MQTT.

  4. Does Safal have MLOps and infrastructure experience?

    Yes. In 2025 Safal led the build of the AIHPC GPU cluster for Kathmandu University’s Department of AI: bare-metal nodes with RTX 3080 Ti GPUs, Kubernetes scheduling, HAProxy load balancing, and Prometheus and Grafana monitoring. In 2022 Safal built and ran a public bare-metal Linux server that hosted the Like Tiffin backend.

  5. What is bala.ai?

    bala.ai is Safal Kumar Shrestha’s independent software label for Android apps, including KU Krishi, TrekSathi and Like Tiffin.

  6. How can I hire or contact Safal Kumar Shrestha?

    Email hello@safalkumarshrestha.com.np or connect on LinkedIn. Safal is based in Chitwan and Kathmandu, Nepal (UTC+5:45), and is open to remote engineering roles, AI/ML internships, research collaborations and app partnerships.

08 Contact

Have a model to ship, a cluster to build or an app to make? I’d like to hear about it.

hello@safalkumarshrestha.com.np

Also at safalkumarshrestha08@gmail.com · Open to research collaborations, AI/ML internships, app partnerships and remote engineering roles.