Neel Mokaria

Software Engineer & AI Researcher

I build search, memory, and multimodal AI systems.

Things I’ve made Hit me up
Scroll Down

A little about me

Hi, I’m Neel.

I like working on systems where the small details matter. Search that returns the right source, memory that can explain where a fact came from, and multimodal agents that give you something to inspect when they are wrong.

That is what pulled me into the work I do now: building search and document infrastructure at Vesara, governed knowledge graphs at MIND Lab, multimodal research at GAMMA Lab, and software for nonprofit teams through Hack4Impact.

Away from the keyboard

SwimmingSquashBooksMusic

A few things I’ve worked on

I make things.

Some with a research lab. Some with a nonprofit. One at a hackathon.

01 / MIND Lab

Agentic
Graph Memory.

MAGG / how a fact gets in01 → 03

Domain-owner agents approve, reject, or revise a proposed fact before it enters the graph.

An illustration of the governance step.

A memory that can explain itself.

I built a Python system that puts extracted facts through review before adding them to a knowledge graph. Each stored fact keeps its evidence and decision history, so it can be inspected and revised later.

On SciERC, governance improved strict triple F1 from 0.060 to 0.080, a 33% relative improvement.

Python / Knowledge graphs / Multi-agent systems / vLLM

Code

02 / A team hackathon project

IronSite.

1st place, UMD 2026 ↗

IronSite result frame highlighting a worker’s ungloved hand as a possible PPE violation.
Result frames from our public repository.

A second look at construction footage.

Our team built a safety dashboard that finds suspected hazards in construction video, checks them with three-pass multimodal verification, and turns the findings into end-of-shift reports.

First place at UMD’s IronSite hackathon, with a $10,000 team prize. The work became a construction safety preprint.

YOLO11 / SAM 3 / Qwen3-VL / FastAPI / Next.js

Code Paper

03 / Hack4Impact-UMD

For All Ages.

Less matching by hand.

With a 14-person team, I built matching and coordination tools for a nonprofit. The pipeline combines embedding similarity, participant preferences, and the Hungarian algorithm for 1,000+ participants.

The admin dashboard saves coordinators roughly 12 weeks per matching cycle.

React / Node.js / PostgreSQL / Firebase / Pinecone

Code

04 / Course project · December 2025

Serverless
Banking.

Following a transaction through AWS.

A backend with six Lambda functions and API Gateway. DynamoDB conditional expressions handle atomic balance updates; SNS and SES handle notifications, and S3 stores generated statements. IAM policies follow least privilege.

Node.js / Lambda / API Gateway / DynamoDB / SNS / SES / S3

Code
A couple of earlier projects Invoices & personal finance

GenAI Bill Automation

Automated vendor-bill processing using OCR, layout-aware models, LLM-based entity extraction, and structured database integration for real-time reporting.

  • Python
  • OCR
  • LangChain
  • Transformers
Code

A.I. Finance Analyser

AI-powered dashboard that categorises bank statement transactions by spending type using Llama 3.1, surfacing actionable financial insights.

Code
What I use Languages, tools, and infrastructure
Languages
Python, TypeScript, JavaScript, Java, C/C++, SQL, Bash
Product & backend
React, Next.js, Node.js, Express, FastAPI, REST, GraphQL
Data & retrieval
PostgreSQL, Prisma, DynamoDB, Firebase, Pinecone, hybrid search, knowledge graphs
AI systems
Multi-agent systems, multimodal models, vLLM, computer vision, reranking
Infrastructure
AWS, Docker, Linux, Git, CI/CD

The people I’ve worked with

Along the way.

July 2026 – August 2026

San Francisco, CA

Vesara

Founding Engineer Intern · Search & Memory Infrastructure

Built search and document infrastructure to help people find the right information across company data.

Work details at Vesara
  • Built a PostgreSQL hybrid search layer combining keyword, vector, and graph retrieval with reciprocal rank fusion and cross-encoder reranking. Improved Recall@5 by 16.6 percentage points and MRR from 0.46 to 0.65 on a 163-query benchmark.
  • Shipped Gmail and Drive ingestion with section-level indexing, permission filtering before text access, and human review for ambiguous classifications. Helped reduce an unfiled-document backlog from 1,561 to zero.
  • Migrated job scheduling to AWS EventBridge Scheduler and Lambda, validating a new event bus against the existing transport on production traffic.
  • PostgreSQL
  • Hybrid Search
  • Reranking
  • AWS
  • EventBridge
  • Lambda

August 2025 – Present

Remote

Hack4Impact-National

Software Engineer · Volunteer Management System

Built GraphQL and REST APIs for a volunteer management system serving 1,000+ members across 16+ university chapters.

Work details at Hack4Impact-National
  • Designed normalized PostgreSQL schemas and type-safe data access with TypeScript, Express, Prisma, Zod validation, and connection pooling.
  • TypeScript
  • Express
  • GraphQL
  • REST
  • Prisma
  • PostgreSQL

September 2025 – Present

College Park, MD

For All Ages × Hack4Impact-UMD

Software Engineer · Team Lead

Built matching and coordination tools with a 14-person cross-functional team, saving coordinators roughly 12 weeks per matching cycle.

Work details at For All Ages × Hack4Impact-UMD
  • Worked with nonprofit stakeholders to scope the product, plan sprints, and review code.
  • Designed a matching pipeline for 1,000+ participants using Pinecone embeddings, structured preference scoring, and the Hungarian algorithm.
  • Built a React admin dashboard with Node.js services, PostgreSQL, and Firebase.
  • React
  • Node.js
  • PostgreSQL
  • Firebase
  • Pinecone
  • Matching Algorithms
View matching project

May 2025 – Present

College Park, MD

MIND Labs, UMIACS · University of Maryland

Research Intern · Knowledge Graph & Agentic Systems Infrastructure

Built governed knowledge graphs in Python, retaining each fact’s source and the decisions behind its admission.

Work details at MIND Labs, UMIACS · University of Maryland
  • Domain-owner agents approve, reject, or revise extracted facts before insertion. Governed fact admission improved strict triple F1 by 33% relative on SciERC.
  • Operated a Linux/vLLM serving stack across two NVIDIA L40S GPUs, reducing batch runtime by 38% through parallel requests and adding embedding-service failover.
  • Python
  • Knowledge Graphs
  • Multi-Agent Systems
  • vLLM
  • Linux
  • Model Serving
Explore graph memory

June 2025 – February 2026

College Park, MD

GAMMA Lab, UMIACS · University of Maryland

Research Intern · Multimodal Agentic AI

Investigated how perception, reasoning, memory, and action interact in multimodal agentic frameworks.

Work details at GAMMA Lab, UMIACS · University of Maryland
  • Analyzed cross-modal integration, benchmark coverage, and scalability to identify research gaps and evaluation limitations.
  • Co-authored a survey under Prof. Dinesh Manocha on the foundations and frontiers of multimodal agentic frameworks.
  • Multimodal AI
  • Agentic Systems
  • Research
  • Evaluation
Read the survey

Research & recognition

Papers & a team win.

Research survey

Foundations and Frontiers of Multimodal Agentic Frameworks

GAMMA Lab, UMIACS, University of Maryland
Research under Prof. Dinesh Manocha

Co-authored a survey examining multimodal agent architectures, cross-modal integration, evaluation, and scalability, with an emphasis on open research gaps.

Paper PDF

arXiv preprint · May 2026

Passive Construction Site Safety Monitoring via Persona-Scaffolded Adversarial Chain-of-Thought VLM Verification

Ananth Sriram, Neel Mokaria, Rajveer Singh

Co-authored a construction-video safety monitoring paper combining detection, segmentation, and three-pass VLM verification to produce evidence-linked, end-of-shift reports.

Paper PDF Code

Keep in touch

Hit me up.

Software engineering, AI infrastructure, or a research collaboration. I’d love to hear what you’re working on.

neelmokaria@gmail.com

Find me on the web