John Acosta

Applied AI & Software Engineering

Now

Final year of computer science at NJIT. Building my foundation in AI to understand the mechanisms behind the technology disrupting the world.

Updated

How I work

  • Understand the why first.
  • Ask if it's feasible, viable, and desirable.
  • Own it end to end.
  • Let feedback shape the idea.

Experience

  1. Production at enterprise scale

    –

    Software Engineer Intern

    Prudential PGIM Technology · Finance

    In regulated finance, production infrastructure has to be secure, repeatable, and simple enough that the next team can trust it.

    Details
    • Developed reusable Terraform modules for AWS Lambda, S3, and KMS to migrate off CloudFormation under enterprise infrastructure-as-code standards spanning multiple business units, cutting build time and infrastructure code for pilot applications.
    • Automated reporting on monthly AWS spend with a React and FastAPI dashboard on S3, replacing a manual process.
    • Led a team of five engineers to second place in PGIM's internal hackathon with an AI-powered Confluence pipeline that detects stale pages, suggests updates, and notifies reviewers, improving the documentation Copilot and Claude MCP rely on.
    • Terraform
    • AWS
    • CloudFormation
    • React
    • FastAPI

  2. From hypothesis to evidence

    –

    Undergraduate Research Assistant

    Lijing Wang Lab, NJIT

    Turning a professor's hypothesis into code means building experiments clean enough that every answer holds up.

    Details
    • Turned a professor's research hypothesis on gradient cancellation into a modular Python framework for rapid iteration, executing 250+ controlled fine-tuning runs across 4 language and vision models on an A100 GPU cluster (6,000+ GPU service units).
    • Audited Hugging Face and PyTorch internals to remove hidden global seed state, building custom seed modules and checkpoints so each source of randomness could be tested in isolation, preventing invalidation of 400+ planned runs.
    • Built Python and Jupyter analysis scripts that turn raw fine-tuning output into structured results reports and visualizations, enabling faster faculty review of experiment outcomes.
    • PyTorch
    • Hugging Face Transformers
    • SLURM
    • Jupyter

  3. Reliable Clinical AI

    –

    Applied AI Research Engineer Intern

    AI Imaging Medicine Lab, NJIT · Healthcare

    A model that reasons like a radiologist still needs tests that prove it before any surgeon sees its output.

    Details
    • Ported a neurosurgical PyQt application to Linux HPC and built an automated QA harness over 100 natural-language prompt suites, capturing UI screenshots and telemetry to isolate failures, driving fixes that raised pass rates from 20% to 100%.
    • Designed a LangChain workflow that breaks one-shot tumor diagnosis into a 5-step radiologist-style reasoning chain (e.g., symmetry, lesion location), each a dedicated model call on BraTS MRI data, delivering a prototype to inform the lab's system design.
    • Python
    • LangChain
    • PyQt
    • Linux
    • HPC

  4. Engineering for research

    –

    Deep Learning Research Intern

    Lijing Wang Lab, NJIT · Public health

    Nobody handed me research experience, so I found my first lab and learned research code needs real engineering discipline.

    Details
    • Parallelized independent forecasting workloads across CPU cores with joblib, cutting experiment runtime by ~80% and enabling faster iteration without HPC resources.
    • Engineered modular Python preprocessing (NumPy, pandas) to clean, trim, and interpolate 10+ multi-location epidemic datasets (~900 weekly observations each), producing consistent train/test splits for rolling-origin forecasting across 5 diseases.
    • Python
    • NumPy
    • pandas
    • joblib

Leadership

  • Member & Role Model · HISPA

    – Present

    Hispanics Inspiring Students' Performance and Achievements

  • Member · Society of Hispanic Professional Engineers, NJIT

    – Present

Projects

  • Intelligent LabComing soon

    An AI layer for research labs built on meeting transcripts, MCP, and agents.