Available for new work

Building ML systems end to end, and checking that they work.

I'm Apoorv Raj — AI Engineer (contract) at Node2.io, building the AI layer of BuildingSync, a PropTech SaaS. I work across FastAPI services, React and TypeScript front ends, and the evaluation that decides whether a model is worth trusting.

Read the flagship case study All work

AI Engineer · Node2.ioCo-author, IEEE paper on image captioningIndia · Remote

About

ML engineering across the stack — research to deployment.

I build the whole path: data, models, evaluation, inference APIs and the product around them. The thread through everything is checking whether a system actually works before claiming it does.

I’m a Computer Science graduate from Chandigarh University (2024) and a co-author of an IEEE conference paper, AI Narratives: Bridging Visual Content and Linguistic Expression, on image captioning with an InceptionV3 CNN encoder and a Transformer decoder trained on COCO.

In 2023 I was an ML Engineering Intern at Quicksilver Technologies, writing Python data-processing scripts, debugging utilities and basic unit tests for software-validation workflows.

Today I’m an AI Engineer (contract) at Node2.io and a founding contributor to BuildingSync, a PropTech SaaS. I ship product features in a Next.js and TypeScript codebase backed by PostgreSQL with Row-Level Security, and I’m building its AI layer: a FastAPI service with an LLM booking assistant that uses LangChain tool calling, and a complaint-prioritization engine.

Outside work, Fraud Radar is the clearest example of how I build: a real-time fraud-scoring system with 1,407 tests and an evaluation showing that its model does not transfer across data generators without retraining, reported rather than hidden.

→

Evaluate before claiming

Every figure names what it was measured on. Fraud Radar froze its benchmark methodology before scoring and published a cross-generator transfer result of PR-AUC 0.0087.

→

Tested, typed boundaries

Pydantic and TypeScript at the edges, strict type checking, and CI gates on every push. Fraud Radar runs 1,407 tests in CI.

→

Decisions written down

Architecture, methodology and limitations recorded next to the code. Fraud Radar alone carries eleven decision records.

Education

B.E. in Computer Science & Engineering

Chandigarh University

Computer Science & Engineering · 2020 – 2024

CGPA: 7.66 / 10
Graduated: 2024

Selected work

One system I can defend in depth, and the work behind it.

Each case study states what was built, how it was evaluated, who did what, and where it falls short.

FlagshipLive demo · 2026

Fraud Radar — Real-time card-fraud detection

Real-time fraud scoring with rules, XGBoost and per-decision SHAP. 1,407 tests, a live demo, and a benchmark that shows where the model fails.

Build
Evaluate
Discover
Report

Read the case study Live demo Source

The Fraud Radar transactions view: scored transactions with decisions, amounts converted to a reporting currency, and filters. — The public demo, reading the frozen snapshot it ships with. There is no hosted backend.

Measured, not claimed

0.9327 PR-AUC

In-distribution

Synthetic held-out fold · 93 frauds in 7,502

0.8653 PR-AUC

Retrained on an independent generator

Sparkov fold · 924 frauds in 277,860

0.0087 PR-AUC

Carried across generators, no retraining

Same Sparkov fold · 0.0033 prevalence

0.7670 PR-AUC

Real anonymised card data, isolated track

ULB seed 42 · 52 frauds in 42,722

3.7 ms

Service-layer scoring, p50

n = 500, developer laptop

1,407 tests

Green in CI

ruff · mypy --strict · tsc · eslint

The same 17-feature pipeline scores 0.9327 on the generator it was trained on and 0.0087 when carried across to an independent one without retraining. Five of its features are constant on the target data, and the operating threshold does not transfer either. That result is published in the repository rather than left out.

Also built

Image Captioning System — CNN + Transformer preview

Research

Demo offline2024 – 2026

Image Captioning System — CNN + Transformer

Co-authored IEEE research rebuilt as a tested Python package and FastAPI service, with an audit of what its BLEU score does and does not show.

PythonTensorFlow / KerasInceptionV3Transformer

Case study

Earlier and collaborative work

Career

From research to applied AI engineering.

A co-authored IEEE paper, an ML engineering internship, and founding-contributor work at Node2.io on a PropTech SaaS and its AI layer.

View full resume

AI Engineer
Node2.io
2026 – Present
Independent contractor · Software Engineer (Founding Contributor) · PropTech SaaS (BuildingSync) · Remote, Canada
- Ship resident- and staff-facing product features across a commercial Next.js (App Router) + TypeScript codebase — UI components through Prisma data-access modules to PostgreSQL (Supabase) under Row-Level Security.
- Built accessibility features for the resident-facing product — large-text and high-contrast display modes — and keep interactive components keyboard-accessible.
- Build the platform’s AI layer as a Python FastAPI service — an LLM booking assistant (LangChain tool calling against a 50+ slot schema, structured response validation, voice interface in development) plus a complaint-prioritization engine.
- Contribute code reviews, Jest unit/integration suites, CI/CD pipelines, Docker deployments, and Linux debugging in Agile sprints, collaborating remotely with a Canada-based team across time zones.
ML Engineering Intern
Quicksilver Technologies Pvt. Ltd.
2023 – 2023
India
- Supported software-validation workflows with Python data-processing scripts, debugging utilities, and basic unit tests.

Research

Multimodal AI, co-authored.

A co-authored IEEE conference paper on image captioning with a CNN visual encoder and a Transformer decoder.

AI Narratives: Bridging Visual Content and Linguistic Expression

Preetam, Sai Chetan Muppalla, Apoorv Raj, Jasneet Chawla

2024

2024 IEEE International Conference on Smart Power Control and Renewable Energy (ICSPCRE)·DOI 10.1109/ICSPCRE62303.2024.10675203

Image captioning with an InceptionV3 CNN visual encoder and a Transformer-based decoder, trained on the COCO dataset.

Read on IEEE Xplore View DOI

Stack

The day-to-day toolkit.

Languages, model frameworks, data pipelines, deployment — the tools I actually reach for, not the resume keywords.

Languages

PythonTypeScriptJavaScriptSQL

Machine Learning

XGBoostscikit-learnSHAPTensorFlow / KeraspandasNumPyLLM tool calling (LangChain)

Backend & Data

FastAPIPydantic v2SQLAlchemy 2.0AlembicPostgreSQLSQLitePrismaREST API design

Frontend

React 19Next.js (App Router)TanStack QueryTailwind CSSAccessible UI

Testing & Quality

pytestJestmypyRuffESLintpre-commitCode review

Delivery

GitHub ActionsDockerVercelHugging Face HubStreamlitGitLinux

Contact

Building something serious?

Open to software, AI and ML engineering roles, and to applied-ML and data-intensive systems work. Email is the fastest way to reach me.

Direct line

Email is the fastest channel.

Architecture, data, deployment — describe the system. The more specific, the easier it is to tell you whether I’m the right engineer for it.

apoorvrajmgr@gmail.com Resume

apoorvrajmgr@gmail.com

/in/apoorv-raj-1a35ba218

GitHub

github.com/apoorvrajdev