PortfolioNewark, New JerseyMMXXVI

Software engineer,
AI systems
end to end

I take AI systems from an ambiguous problem to something deployed and used. Most of my work starts as a vague question — can someone who doesn't write SQL just ask this database? — and ends as a system somebody runs.

Author: Vineet Vora
Programme: MS Artificial Intelligence, NJIT
Conferred: June 2027
Status: Open to Forward Deployed Engineer roles

Read the work Get in touch

Abstract

I like the part of the job where nobody has decided what to build yet.

That usually means owning the whole line: scoping what's actually needed, building the pipeline, standing up the infrastructure it runs on, and then explaining the result to people who don't care what a transformer is.

A model isn't finished when it converges. It's finished when someone who didn't build it can run it, trust it, and get an answer under real constraints.

End-to-end ownership. Ambiguous problems. Systems that run in someone else's environment. Shipping over polishing. Knowing when to cut scope.

Selected work

Ask your database in English

Natural language to SQL over enterprise databases the model has never seen.

What if someone who doesn't write SQL could just ask? The system turns plain-English business questions into working SQL against unfamiliar multi-table schemas — schema linking, candidate generation, targeted decomposition, self-repair, consensus vote.

It runs on open-source models only. No proprietary API, which means it can be deployed inside a company's own environment with no data leaving it — the difference between a demo and something a regulated customer can actually run.

Results on the BIRD benchmark
Execution accuracy	70.2%
Benchmark questions	1,534
Distinct databases	11
Proprietary APIs	0

Thesis in progress. Repository not yet public.

Forecasting geomagnetic storms

A frozen encoder beats LoRA on a 366M-parameter Vision Transformer pretrained on NASA solar imagery.

Can you see a storm seventy-two hours out? Adapted Surya to forecast the interplanetary magnetic field conditions utilities and satellite operators watch before a geomagnetic storm — comparing a frozen encoder against LoRA on 13-channel AIA EUV and HMI magnetogram inputs at 512×512.

Forecast performance
RMSE	3.39 nT
Cross-validation folds succeeded	96.5%
Efficiency vs. LoRA	3.5× more efficient

github.com/VineetV2/imf-bz-forecasting-surya

118.9 million flight records

Distributed analytics on a 6-node Hadoop cluster, stood up from bare EC2 instances.

Twenty-two years of US flight data, processed with custom MapReduce jobs and then tuned until every map task read local data rather than pulling across the network.

Cluster performance
Records processed	118.9M
Dataset size	11.66 GB
Data-local map execution	100%
Speedup under scaling	2.73×

github.com/VineetV2/hadoop-flight-data-analysis

Peon — memory for AI coding agents

A local-first MCP server, shipped and used daily.

Coding agents forget everything between sessions. Peon gives them persistent project memory: capture as you work, LLM consolidation, hybrid retrieval, a daily self-audit. It ships as an MCP server, so it drops into any client rather than locking to one.

github.com/VineetV2/peon-mem

This document

You're reading it.

Hand-written HTML, CSS and JavaScript. No framework, no build step, no dependencies beyond three typefaces.

github.com/VineetV2/portfolio

Face-based attendance automation

No manual roll call — attendance logged straight from the camera feed.

A live classroom camera feed runs through YOLO for face detection, then FaceNet for identity matching, logging each recognized student straight to an Excel sheet.

github.com/VineetV2/Face-Based-Attendance-Automation

Methods

Technical stack by domain
Languages	Python · TypeScript · Java · JavaScript · SQL
AI & ML	PyTorch · HuggingFace · vLLM · LoRA / PEFT · TensorFlow · scikit-learn · OpenCV
Data & infrastructure	Hadoop · MapReduce · HDFS · YARN · AWS EC2 · SLURM · Docker · CUDA
Databases	PostgreSQL · MySQL · MongoDB · SQLite
Web	React · Node.js · Cloudflare Workers
Delivery	Requirements scoping · Design Thinking (IBM) · Technical writing · Git · CI/CD

Published

Preventing Wildfires in Energy Transmission by Automatic Power Line Defects Detection Using Machine Learning and AI

A computer vision framework for finding power line defects and vegetation encroachment in UAV imagery — moving utilities from reactive repair to preventive maintenance before a fault becomes a fire.

The proof of concept detects falling wires using Canny edge detection and a Probabilistic Hough Transform. CNN, YOLO and transformer architectures are the roadmap toward real-time inspection at scale.

ieeexplore.ieee.org/document/11380637

Record

Education
2027	New Jersey Institute of Technology MS, Artificial Intelligence · GPA 3.83 · Newark, NJ Advanced NLP · Computer Vision · Information Retrieval · ML for Data Science · Big Data
2025	Vellore Institute of Technology — AP BTech, Computer Science · India

Certifications

Contact

Building something where the hard part is figuring out what to build? That's the part I want.