Open to engineering roles

Neeraj Bala

Full-Stack AI Engineer · M.Sc. Computer Science · Stockholm

See my work Résumé Stockholm, Sweden

Selected work

VALI Studio — e-commerce platform

A production storefront for a Swedish beauty store with ~1,000 in-stock products from 22+ brands. AI search and a shade finder up front; a 16-tab admin with an AI invoice pipeline that cross-checks supplier invoices with two models. Stripe checkout on Supabase — built solo, from schema to deploy.

React 19StripeSupabaseAI invoices

SkyYard — yard management system

Tracks every truck from gate arrival to dock to exit across six role-based interfaces — guard, hostler, admin, superadmin, carrier and driver. Live Mapbox map, drag-and-drop dock scheduling, kiosk check-in and Socket.IO updates. 176/176 tests passing.

React 19MapboxSocket.IOExpress

CareerOS — AI job search platform

EU/Sweden-focused job search automation: discovers roles from 350+ sources, scores them against your profile with LLMs, writes tailored resumes and cover letters, then auto-applies through real ATS forms with a human-in-the-loop confirm step.

Next.js 15FastAPICeleryPlaywright

Job Tracker

Upload a JD and CVs; AI extracts company, role and location, and a cover-letter tool fills a LaTeX template, compiles it to PDF and saves the application only after approval. Flask, Neon Postgres and Google Drive, deployed on Render.

FlaskPostgreSQLLLMLaTeX

Experience

Ericsson logo2025

Ericsson

DevOps & Test Developer Intern

Built GitLab CI/CD pipelines and automated hardware-dependent software testing with Docker, C# and LabVIEW.

U-Blox logo2022—23

U-Blox

Machine Learning Intern

Evaluated anomaly detection models on routing data using Python, Power BI and Tableau.

Luday SE logo2024

Luday SE

Full Stack Developer Intern

Helped ship a production MVP; API improvements reduced data latency by 25% and review practices reduced post-release bugs by 30%.

NB2026—Now

Freelance

Full-Stack AI Engineer

Shipped an e-commerce platform and AI inventory workflow; now developing real-time yard management software.

Professional proof

“Neeraj is a hard working guy. He's honest about what he can and can't and would do anything to find answers to things he didn't know.”

Hari VigneswaranSenior Product Manager, IoT & Cybersecurity · Managed Neeraj directly · January 19, 2026

Education & technical depth

Blekinge Institute of Technology, Sweden logo

M.Sc. Computer Science

Blekinge Institute of Technology, Sweden

Thesis: Adapting MotionBERT for Sports Motion Analysis.

2024—2026
Blekinge Institute of Technology, Sweden logo

B.Tech. Computer Science

Blekinge Institute of Technology, Sweden

2022—2024
JNTU Kakinada, India logo

B.Tech. Computer Science & Engineering

JNTU Kakinada, India

2018—2022

Technical skills

AI engineering

  • Python
  • AI agents
  • RAG
  • Vector databases
  • FastAPI
CareerOS · AI search & automation

Full-stack development

  • TypeScript
  • React
  • Next.js
  • Express
  • Socket.IO
VALI Studio · SkyYard

Cloud & delivery

  • AWS
  • Azure
  • Docker
  • Kubernetes
  • Terraform
  • GitLab CI/CD
Ericsson · CI/CD & test automation

Data & quality

  • SQL
  • PostgreSQL
  • MySQL
  • Supabase
  • Playwright
Job Tracker · CareerOS

Also worked with · C++ · .NET · Angular · Flask · Celery

Research & publications

Pose reconstruction error: zero-shot MotionBERT 0.9670, random-init DSTformer 0.9064, biGRU 0.0747, fine-tuned MotionBERT 0.0187.
Reconstruction comparison, redrawn from thesis Table 5.3. MPJPE is reported in torso-normalised units; lower is better.

Adapting MotionBERT for Sports Motion Analysis: Extraction, Comparison, and Generation of Continuous Poses

Adapted a pretrained motion transformer to 654 clips across five sports. Built a skeleton-remapping pipeline and evaluated pose accuracy, movement continuity and biomechanical indicators.

Key finding

98.1% lower pose error than the zero-shot baseline; 75.6% lower than biGRU. High-frequency jitter remains a limitation, and the system is not a medical diagnostic tool.

MotionBERT · SportsPose · 3D pose estimation · Temporal evaluation

Neeraj Bala · Supervisor: Shahryar Eivazzadeh · 106 pages

Iris recognition accuracy: CNN 98.7%, SVM 86%, SVM with Hamming distance 80%.
Algorithm comparison, redrawn from thesis Table 5.3. These are the report’s experimental results, not a general benchmark.

Comparative Analysis of Machine Learning Algorithms for Biometric Iris Recognition Systems

Compared CNN, SVM and SVM with Hamming distance for biometric iris recognition, combining a literature review with experiments on iris images and a 60/20/20 train–test–validation split.

Key finding

CNN achieved 98.7% accuracy, compared with 86% for SVM and 80% for SVM with Hamming distance, on the evaluated dataset.

Computer vision · CNN · SVM · Feature extraction

Neeraj Bala & Vishnu Kiran Dabbara · Supervisor: Suejb Memeti

Certifications

Looking for an engineer who can connect AI research to real products?

neerajnani211@gmail.com