Research AI Engineer · Machine Learning

Nhoojah
Maharjan

Machine learning, reinforcement learning, and world models —
building agentic systems that hold up in the real world.

Nhoojah Maharjan

Hi, I'm Nhoojah

Computer Science undergraduate at the University of Southern Mississippi (GPA: 4.0), working across machine learning research and the data infrastructure that research depends on. I study reinforcement learning and world models, and publish on segmentation and machine unlearning. Alongside that I build production systems — NOAA oceanographic data platforms, public-health data pipelines, and LLM agent surfaces.

Research Interests

Machine learning, reinforcement learning, and world models — with a particular interest in causal reasoning, sample-efficient exploration, and deployable agentic systems that support reliable, real-world decision-making. Open to new directions across ML and autonomous systems.

University of Southern Mississippi
B.S. Computer Science · Honors Scholar · Minors in Economics & Data Analysis and Mathematics · GPA 4.0
Aug 2024 – May 2028 (expected) · Hattiesburg, MS
Coursework: Data Structures & Algorithms · Artificial Intelligence · Linear Algebra · Probability & Mathematical Statistics · Discrete Mathematics · Computer Systems · Computer Networks · Database Concepts · Secure Software Development · Cyber Law & Ethics
01
Reinforcement Learning & World Models

Model-based RL for task-agnostic exploration — action-rooted causal affordance models, compositional capability graphs over open-world environments, and planning in latent imagination with uncertainty-gated grounding.

02
Foundation Models & Unlearning

Cross-domain benchmarking of vision-language models against supervised baselines, and exact machine unlearning that lets fine-tuned models forget specific training data without a full retrain.

03
LLM Agents & Tooling

Agentic workflows, RAG, and Model Context Protocol surfaces that expose large scientific catalogs to LLM clients, backed by GCP-hosted inference.

04
Data & Systems Engineering

Multi-source ingestion and indexing pipelines, typed RPC services in Go, high-throughput backends, and Terraform-managed cloud infrastructure across GCP, AWS, and Azure.

  • Research AI Engineer
    Nov 2025 – Present
    Institute for Advanced Analytics and Society (IAAS) · NOAA-funded · Hattiesburg, MS
    • Contributing to AQUAVIEW, a production cloud-native platform giving researchers and downstream pipelines programmatic access to over a century of NOAA oceanographic and Earth-observation data.
    • Shipped AQUAVIEW's production MCP server on Cloud Run, giving AI agents identity-aware access to the platform's catalog and data.
    • Built a LangGraph data agent on Gemini that keeps restricted data out of prompts and keeps cost bounded with usage budgets and tracing.
    • Building and maintaining the Python data library and REST API that expose the catalog to human researchers and to LLM/agent clients.
    • Supporting the multi-source ingestion and indexing pipeline (Dagster + STAC/SFEOS on Elasticsearch), unifying 40+ heterogeneous NOAA sources into a single searchable catalog covering 125+ years of ocean data.
    • Prototyped LLM-based tagging and classification for platform metadata.
  • Research AI Engineer
    Sep 2026 – Present
    Client: USM Office of Research Administration · IAAS · Led technical discovery with the Associate VP
    • Covered the 39% of test obligations found only in incorporated regulations by building PI Award Kickoff, a RAG-like LLM pipeline on Cloud Run that turns a federal award and its regulations, like NIH's 427-page policy, into a cited checklist in ~5 min.
    • Kept hallucinated quotes from shipping by requiring every checklist item to quote the exact text the model saw, enforced by a CI evaluation gate, and logged tokens, latency, and cost per call ($2.38 on a 1M-token NIH sample award).
    • Built Researcher Profile Builder for federal research-security review, integrating 5 scholarly and award APIs with PII redaction: all 3,396 claims link to hash-checked sources, and it caught a record where 21 of 49 works belonged to namesakes.
  • Undergraduate Researcher
    Apr 2025 – Present
    Advised by Dr. Rabab Abdelfattah · AIMS Lab, University of Southern Mississippi
    • Co-authored the first comprehensive benchmark of deep-learning architectures for cross-domain thin-object segmentation — a 5-family taxonomy (CNNs, Transformers, GANs, classical, foundation models) evaluated across biomedical, infrastructure, and remote-sensing datasets (IEEE Internet of Things Journal, 2026).
    • Stress-tested vision-language foundation models (SAM, DIS-SAM, CLIPSeg, GSVA) against supervised baselines under zero-shot and prompt-engineered conditions, quantifying up to a 13% absolute precision increase on infrastructure benchmarks.
    • First-authoring SEU (Siloed Exact Machine Unlearning) — an architecture that lets fine-tuned models exactly forget specific training data by pairing a frozen foundation encoder with isolated adapter heads, making deletion a targeted reset rather than a full retrain (~8,700× faster than exact retraining while matching a from-scratch model).
  • Undergraduate Researcher
    Jun 2026 – Present
    Advised by Dr. Alen Hajnal · Perception, Action & Cognition Lab, USM
    • Developing a model-based reinforcement learning (MBRL) framework for task-agnostic exploration that replaces standard pixel-prediction dynamics with an action-rooted Causal Affordance World Model.
    • Designing compositional capability graphs over open-world environments (Minecraft, Crafter) where each edge encodes an agent-centric causal intervention; planning happens in latent imagination and drops into the real environment only when the model ensemble is uncertain about a specific affordance edge.
    • Formalizing self-termination via graph completeness — the framework inherently prunes the classic "Noisy TV" pathology, since uncontrollable dynamics produce no stable causal links.
  • Undergraduate Researcher
    Fall 2024
    Advised by Dr. Partha Sengupta · University of Southern Mississippi
    • Investigated Virtual Reality Exposure Therapy (VRET) for PTSD, phobias, and anxiety, weighing clinical effect sizes against biometric-data privacy, dissociation risk, and the security gaps that come with immersive data collection.
    • Presented findings as a poster at the 89th Mississippi Academy of Sciences Annual Conference, Biloxi Gulf Coast Convention Center.
  • Data Engineering Intern
    May – Aug 2026
    Gulf South Center for Community-Engaged Health Research and Innovation · Hattiesburg, MS
    • Cut retrieval latency 70% for researchers, hospitals, and labs by designing the SQL schema the team adopted and ETL/ELT pipelines unifying structured CDC and Census/SVI data with 500+ GB of unstructured MRI imaging in all 82 Mississippi counties.
    • Designed multi-layer APIs and metadata cataloging so researchers can pull cross-source datasets without dealing with schema differences or curation overhead.
    • Implemented versioned dataset snapshots so downstream analyses stay reproducible against precisely dated slices of federal data.
  • Software Engineering Intern
    Jan – May 2026
    Optimal Answers LLC · AI decision-optimization software · Gulfport, MS
    • Cut customer onboarding from 3 weeks to 4 days by meeting directly with customer engineering teams to scope requirements and building automated ETL integrations connecting legacy ERP and spreadsheet data to the company's optimization models.
    • Architected prescriptive analytics models using Dataform (Optimal Modeling Language) on Azure to solve Linear and Integer Programming problems, modeling qualification criteria, solution constraints, and reduced-cost marginal-value outputs across procurement, portfolio, and supply chain pipelines.
    • Modeled large economic datasets from the Gulf Coast Business Council and Restoration Fund to optimize grant selection and resource allocation, defining multi-period constraint hierarchies and attribute logic to maximize goal achievement under strict bounds.
  • Software Engineer
    May – Aug 2025
    Dabuli (Nonprofit, 501(c)(3)) · Seattle, WA
    • Cut p95 API latency 84% (1.9 s to 310 ms) with Redis caching in a TypeScript/Node and PostgreSQL API.
    • Built the team's first CI/CD with staged, auto-rollback deploys, cutting releases from 40 minutes to under 5.
    • Instrumented OpenTelemetry distributed tracing and automated recovery pipelines, reducing MTTR from hours to under 10 minutes.

Languages

Python · TypeScript · Go · Java · C/C++ · SQL

ML & AI

PyTorch · HuggingFace Transformers · JAX · PEFT/LoRA/QLoRA fine-tuning · DeepSpeed/FSDP distributed training · model quantization (ONNX, GPTQ) · LM-Eval-Harness · WeightWatcher · machine unlearning

LLMs & Agents

LangGraph · Google ADK · CrewAI · agentic workflows · Model Context Protocol (MCP) · RAG · vector databases (pgvector, Pinecone, Weaviate) · vLLM/TGI inference · prompt engineering · foundation-model APIs (Claude, GPT, Gemini)

Evals & Observability

LLM-as-a-judge · golden datasets · guardrails · granular tracing · tokens/sec · cost per request

Systems & Web

FastAPI · Spring Boot · Node.js/Express · Next.js · React · ConnectRPC/gRPC · GraphQL · Protobuf · WebSockets · Apache Kafka

Cloud & DevOps

GCP (Vertex AI, Gemini, Cloud Run, Cloud Build, Cloud Storage, Firestore, Cloud Trace, IAM, GKE, BigQuery) · AWS (ECS, Fargate, S3) · Azure · Docker · Kubernetes · Terraform · GitHub Actions · OpenTelemetry (Grafana, Tempo, Loki)

Data

PostgreSQL · MongoDB · Elasticsearch · Redis · Firebase · Supabase · Apache Airflow · dbt

  1. From Classical Pipelines to Promptable Foundation Models: A Cross-Domain Study of Thin-Object Segmentation
    A. Hossain, N. Maharjan, M. Hasan, R. Abdelfattah, M. Ezz-Eldin, X. Wang, M. M. Fouda, K. Abdelfattah
    IEEE Internet of Things Journal 2026
  2. SEU: Siloed Exact Machine Unlearning for Fine-Tuned Models
    N. Maharjan, B. Khadka, R. Abdelfattah
    Under Review
  3. Exploring the Promise of Virtual Reality Exposure Therapy (VRET) While Navigating Security Risks in a Digital Age: Examining the Positive and Negative Effects of VR
    N. Maharjan, B. Khadka, A. Kharal
    Poster 89th MAS Annual Conference, Biloxi Gulf Coast Convention Center, MS 2024
Technology for Health Award
College of Nursing and Health Professions, University of Southern Mississippi · Summer 2026 — competitive research grant funding technology work in community health.
NSF ASPIRE Grant
National Science Foundation.
Hatchery Checkpoint Award
$500 seed funding awarded for a project pitch, University of Southern Mississippi.
Honors Keystone Scholar
University of Southern Mississippi.
President's List
Awarded every semester of enrollment, University of Southern Mississippi.
Academic Excellence Scholarship
Full tuition, University of Southern Mississippi.
Enterprise Data Toolkit

Agentic research platform for NOAA, serving 125 users. Hierarchical multi-agent research — a planner splits each request into 2–4 angles that parallel researcher agents work in one cloud sandbox via an 11-tool MCP server — produces cited reports (59–85 sources) in a median 231 s. Its Go sandbox orchestrator, deployed on 7 Terraform-managed Cloud Run services, ran 1,000 round trips with 0 errors and 12 concurrent sessions with no cross-talk.

GoMCPTerraformCloud RunMulti-agent
Research.it

Open-source research assistant built on hybrid RAG: BM25 and vector search over pgvector, fused with Reciprocal Rank Fusion, raised top-1 retrieval accuracy from 57% (embeddings alone) to 74% on 35 labeled questions. A LangGraph multi-agent system pairs hierarchical delegation (a supervisor over parallel ReAct specialists) with a critic loop for self-reflection, plus a reusable 6-tool, token-authenticated MCP server consumed by a Google ADK agent. Requiring validated citations raised grounded answers from 62% to 100% on a 26-question set, and OpenTelemetry traces, exportable to Cloud Trace, meter tokens, tokens/sec, latency, and cost per LLM call.

PythonFastAPILangGraphpgvectorMCPGoogle ADK
EGI Workbench

Research decision-support tool for the Gulf South Center, built across both rounds of its 2026 Hack-a-thon, that turns the Mississippi Health Equity Gap Index into an interactive workbench over all 82 counties. Program officers drill into any county's audit trail, compare counties side by side, and build cohorts exported as PDF reports with methodology citations and the source SQL. A Reweight Lab re-ranks every county under alternative weightings in ~200 ms to test whether a recommendation holds, and "Ask the EGI" turns plain-English questions into SQL with Llama 3.3 70B, showing the query before it runs against a read-only database.

Next.jsTypeScriptSQLiteLlama 3.3LeafletVercel
Vinaya Journal

Privacy-first, offline-capable desktop journaling app with a local LLM + RAG pipeline (Ollama), mood classification, and semantic search across entries. Electron/React frontend, Spring Boot REST backend on SQLite, and a FastAPI microservice for embeddings and vector search — 35+ active users, 150+ GitHub stars across Windows and Linux.

ElectronReactSpring BootFastAPIOllamaSQLite
Learnova

AI-powered educational app with personalized course generation, adaptive quizzes, and smart flashcards powered by Gemini API and Firebase real-time sync.

React NativeExpoFirebaseGemini
HiveFind

Find hackathon teammates by skill match. AI-assisted task generation, real-time chat with AI suggestions, and invitation-based team formation.

React NativeExpoFirebaseJavaScript

Let's work together

Open to research collaborations, engineering roles, and interesting projects. Reach out anytime.

nhoojahm@gmail.com Hattiesburg, MS