Looking for AI/ML engineer roles. Open to relocation

Focus: Multi-Agent LLM Orchestration & Production RAG

AI/ML Engineer
Shipping Production Systems

I'm an AI/ML engineer with five years of experience, mostly LLM agents, RAG on AWS, and computer vision on edge hardware. I care most about what happens after the demo works once: running it on real documents and real devices. M.S. in Data Science from UMBC. Co-author on three papers.

Currently building:

View Projects Resume harshrofff@gmail.com

QC First-Pass
MARS, preliminary 15 runs

Review Time Cut
Lambda + Bedrock automation

Docs in a Production RAG
pgvector · Claude on Bedrock

Publications
SPIE + IEEE + Elsevier

// about

Who I Am

I build AI systems and get them running in real conditions. At VITG that meant a compliance RAG system on AWS Bedrock. In research it meant multi-agent pipelines and computer vision on Jetson boards and a Spot robot.

I have an MS in Data Science from UMBC (3.8 GPA). I did research at UMBC's CARDS lab, funded by the US Army Research Lab, and co-authored three papers (SPIE, IEEE ITU Kaleidoscope, Elsevier JAFR). I hold the AWS Certified Machine Learning Engineer – Associate certification.

The gap between "demo" and "deployed" is where most of the engineering happens. That's the part I like.

MS Data Science

UMBC, GPA 3.8/4.0

AWS ML Certified

Active through Apr 2028

3 Papers Co-Authored

SPIE + IEEE ITU + Elsevier JAFR

Army Research Lab

Funded Research · UMBC CARDS Lab

// skills

Technical Stack

LLM Stack

Multi-Agent Systems RAG Architectures Neuro-Symbolic AI LLM Fine-Tuning (LoRA/QLoRA) LangGraph PydanticAI LangChain LlamaIndex vLLM Prompt Engineering LLM-as-Judge Claude 3.5 Gemini 2.0 Llama 3.1 OpenAI API Ollama

Core AI / ML

PyTorch TensorFlow Vision Transformers (ViT) DINOv2 YOLOv8 Computer Vision Transformer Architectures OpenCV Semantic Search Embedding Pipelines A/B Testing Anomaly Detection Time Series

Cloud & MLOps

AWS Bedrock SageMaker Lambda EC2 S3 Textract CUDA MLflow Weights & Biases Docker GitHub Actions Production Observability

Data Engineering

pgvector Semantic Search Pandas NumPy Apache Spark Kafka PostgreSQL Redis Pinecone ETL/ELT Pipelines Feature Engineering Embedding Pipelines Data Modeling

Software & Deploy

Python FastAPI Flask REST APIs SQL Git Linux/Bash Supabase APScheduler Streamlit NVIDIA Jetson Edge Computing Raspberry Pi

// featured

Production AI Systems

Numbers are from my own runs. Where a test was small or I measured it myself, it says so.

MULTI-AGENT LLM · OPEN SOURCE

MARS: Multi-Agent Research System

A research assistant made of 11 LLM agents in a LangGraph state machine. It writes reports where every citation points to a retrieved source chunk. A QC agent scores each draft and sends weak ones back for another search, and the retrieval tools run through an MCP server.

Problem Research agents hallucinate sources and stop at the first draft, with no grounding and no self-correction.

Approach ReAct researcher (MCP tools) → structured analyst → 3 parallel critics → synthesizer → QC gate that retries with a new search strategy when the score is too low.

Outcome 13 of 15 test runs passed QC on the first try (preliminary). No broken citation tags, which checks that the source exists, not that it supports the claim.

Researcher (MCP) Analyst Parallel Critics ×3 Synthesizer QC Gate (retry)

13/15 QC first-pass 0 broken citation tags 11 agents

LangGraph Gemini 2.5 Flash PydanticAI MCP Streamlit

View Code

EDGE AI · VOICE + VISION

VEST: Voice Assistant on a Jetson

An offline-first voice and vision assistant for a Jetson worn on a vest. You ask a question out loud, it finds the matching procedure out of 100 SOPs, looks at the camera frame, and speaks the answer.

Problem A field assistant that has to work without a reliable connection to a cloud API.

Approach Vosk speech-to-text, CLIP + ChromaDB retrieval over the SOPs, Qwen3-VL through Ollama for the camera frame, Piper for speech. If the camera, model, battery or temperature fails, it says so out loud and drops to a simpler mode.

Outcome About 4.1 s from spoken question to spoken answer on an AGX Orin 64GB with a warm model, and about 13.8 s on an Orin NX 16GB with a smaller model. Both are my own measurements.

Qwen3-VL Ollama CLIP · ChromaDB NVIDIA Jetson Vosk · Piper

NEURO-SYMBOLIC AI

Bio-Oracle

A neuro-symbolic agent for drug-discovery screening. Cellpose segments the cells, then a PydanticAI agent (Gemini) answers questions by calling statistical tools over the extracted features, so each answer can be traced.

Approach Cellpose segmentation → 30+ morphological features → robust Z-score normalization → PydanticAI + Gemini agent.

Outcome 98% outlier precision flagging Taxol-induced actin changes on the BBBC021 dataset. Runs in Docker.

PydanticAI Gemini Cellpose Docker Apple MPS

View Code

RAG · SEMANTIC SEARCH · PROD

Enterprise RAG Platform

Semantic search over 2,500+ regulatory and policy documents, built at VITG for questions where a wrong answer matters. Retrieval on pgvector, answers from Claude on Bedrock.

Problem Staff searched 2,500+ unstructured regulatory PDFs by hand.

Approach AWS Textract → pgvector semantic indexing → Claude 3.5 (Bedrock) → SOC2-aligned audit logging.

Outcome Live in production at VITG. It is company code, so there is no public repo.

AWS Bedrock pgvector Semantic Search EC2 Streamlit

Proprietary (VITG Corp)

CV · PUBLISHED

FilletCam AI

Problem Judging fish fillet quality by eye is slow and inconsistent.

Approach YOLOv8 + OpenCV models for fillet quality analysis. I co-authored the paper.

Outcome Published in Elsevier JAFR 2024.

YOLOv8 OpenCV

Elsevier JAFR

EDGE AI · PUBLISHED

Mosquito Wing Beat Classification

Problem Identify mosquito species in the field, where there is no network or GPU.

Approach Wing-beat sound classifier for a tiny embedded board, built with Edge Impulse.

Outcome Published at IEEE ITU Kaleidoscope 2021.

ARM Cortex-M TinyML Python

IEEE ITU

INFERENCE · BENCHMARK

vLLM Inference Benchmark

Problem Most of my LLM work runs on hosted inference, so I measured a self-hosted engine myself.

Approach vLLM against plain HuggingFace serving, Qwen2.5-3B on a T4, batch sizes 1–32.

Outcome 1.6–2.2x the throughput at batch 4 and up, with matched output tokens. My first run was skewed by unseeded sampling; I found it and fixed it.

vLLM PyTorch CUDA

View Code

AI PLATFORM · LIVE DEMO

Silicon Oracle

A stock research and paper-trading app. Each user brings their own API keys, so rate limits never hit everyone at once. It has a 15-factor score, Gemini analysis, email alerts and Alpaca paper trading.

Problem Good AI stock analysis usually sits behind a subscription or a shared rate limit.

Approach BYOK Flask platform: Finnhub market data → 15-factor Oracle Score → Gemini 2.0 Flash AI analysis → Alpaca paper trading → automated hourly email intelligence alerts.

Outcome Live multi-user app with 70 commits and CI, hosted free on Render. The free host can take a while to wake up.

Flask Gemini 2.0 Flash PostgreSQL Alpaca API Supabase Tailwind CSS APScheduler

Live Demo, No Login View Code

// projects

More Work

AI Video

RECEIPTS

A video agent that cites its clips. Newer footage marks older claims as superseded, and an answer is flagged stale when the footage stops. I built it solo at the Real-Time Video Agents Hack in New York, October 2026. On a 25-question synthetic eval it got 25 right against 15 for a plain RAG baseline.

Python · SQLite · NVIDIA Cosmos Reason

AI OpenAI

AirbnbLens: Listing Search by Style

AirbnbLens: search 62,329 Airbnb listings by photo or by describing a style, using ResNet152 embeddings in ChromaDB.

Python · FastAPI · ChromaDB · ResNet152 · Next.js

Data GeoSpatial

Infrastructure Risk Assessment

Geospatial risk overlays for transportation networks, on an interactive Leaflet map with a PostgreSQL/PostGIS backend.

Flask · PostgreSQL · PostGIS · Leaflet

AI Gemini

SeatSniper

A crowdsourced study-spot finder for UMBC commuters. Availability syncs live through Firebase, and a Gemini planner fits study blocks around your class schedule.

React · Vite · Firebase RTDB · Gemini 2.0 Flash

AI Voice

PelloHorter

A proof of concept for an AI phone agent on Twilio and Gemini. It answers calls and handles plan questions from a local knowledge base.

Python · FastAPI · Twilio · Gemini

CV Edge

Gesture-Based Control System

Gesture teleop for a Jetbot on a Jetson Nano, built from NVIDIA's Jetbot template.

Python · OpenCV · NVIDIA Jetson

// experience

Where I've Worked

Mar 2023 – Present (Concurrent)

Forward Deployed Engineer

UMBC Center for Real-time Distributed Sensing and Autonomy

Built the Spot SDK layer and the HoloLens 2 control path for our multi-robot work, so one person in a headset can drive up to three Spots. A teammate built the gesture recognition. Ran live demos at UMBC's GRIT-X 2024 showcase. Demo ↗
Contributed to a multimodal system on Spot (RGB, thermal and audio with a fine-tuned vision-language model) under US Army Research Lab funding. I co-authored the SPIE 2026 paper about it (third of five authors).
Built a voice and vision assistant with one agent per capability, running open-weight models through Ollama on Jetson, and fine-tuned models with LoRA/QLoRA for intent classification.

Spot SDK HoloLens 2 Ollama NVIDIA Jetson PyTorch

Jun 2024 – Jun 2026

AI Engineer

VITG Corp., Halethorpe MD

Built a semantic RAG platform over 2,500+ regulatory PDFs with pgvector and Claude 3.5 on AWS Bedrock, including permission-gated MCP tools and SOC2-aligned audit logging
Built a serverless automation (Lambda, EventBridge, Bedrock) for review workflows that cut review time by 60%
Ran the infrastructure on Terraform, Docker, FastAPI and GitHub Actions, at 99.9% uptime over two years
Evaluated retrieval with an LLM-as-judge setup in MLflow

AWS Bedrock pgvector Lambda Terraform MLflow

Aug 2023 – Dec 2023

Data Science Intern, Computer Vision

The Conservation Fund, Shepherdstown WV

Built YOLOv8 + OpenCV models for fish fillet quality analysis
Co-authored the resulting paper in the Journal of Agriculture and Food Research (Elsevier, 2024)

YOLOv8 OpenCV Computer Vision

May 2020 – Jun 2022

Machine Learning Engineer

WeHear Innovations Pvt. Ltd., Ahmedabad India

Hearing-health ML for bone-conduction audio devices on an IoT platform
Built time-series feature pipelines for ambient noise, audio exposure and frequency bands
Worked with the firmware and mobile teams on real-time audio processing and small on-device models

Python Time Series Audio Processing Mobile Integration

Master of Science, Data Science

University of Maryland Baltimore County (UMBC)

Aug 2022 – May 2024 • GPA: 3.8 / 4.0

ML · Deep Learning · NLP · Computer Vision · Big Data Systems · Statistical Analysis

// research

Publications & Credentials

Peer-Reviewed Publications

SPIE · Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications VIII · 2026

Multimodal AI for dismounted soldier-robot teaming and context-aware threat response

Gurunathan, Kundurthi, Shroff (third of five), Raza, Gangopadhyay

A multimodal system for soldier-robot teaming that combines RGB, thermal and audio on a Boston Dynamics Spot, under U.S. Army Research Lab funding. I am a co-author.

DOI: 10.1117/12.3093997

Elsevier · Journal of Agriculture and Food Research · 2024

FilletCam AI: A handheld tool for precise fillet color profiling of Atlantic salmon and rainbow trout

Ranjan, Shroff, et al.

A handheld tool for fillet color profiling. I co-authored it during my internship at The Conservation Fund.

DOI: 10.1016/j.jafr.2024.101461

IEEE · ITU Kaleidoscope Conference · 2021

Identification of Deadliest Mosquitoes Using Wing Beats Sound Classification on Tiny Embedded System Using Machine Learning and Edge Impulse Platform

Trivedi, Shroff

Mosquito species identification from wing-beat sound, running on a small embedded board.

DOI: 10.23919/ITUK53220.2021.9662116

Certifications

Certified Machine Learning Engineer – Associate

Amazon Web Services · Active through Apr 2028

Online Presence

Google Scholar

Publications & Citations

// contact

Get in Touch

Looking for AI/ML engineer roles. Open to relocation.
Email is the fastest way to reach me.

✓ Copied to clipboard

LinkedIn GitHub Scholar

AI/ML Engineer Shipping Production Systems

Who I Am

Technical Stack

LLM Stack

Core AI / ML

Cloud & MLOps

Data Engineering

Software & Deploy

Production AI Systems

MARS: Multi-Agent Research System

VEST: Voice Assistant on a Jetson

Bio-Oracle

Enterprise RAG Platform

FilletCam AI

Mosquito Wing Beat Classification

vLLM Inference Benchmark

Silicon Oracle

More Work

RECEIPTS

AirbnbLens: Listing Search by Style

Infrastructure Risk Assessment

SeatSniper

PelloHorter

Gesture-Based Control System

Where I've Worked

Forward Deployed Engineer

AI Engineer

Data Science Intern, Computer Vision

Machine Learning Engineer

Publications & Credentials

Peer-Reviewed Publications

Multimodal AI for dismounted soldier-robot teaming and context-aware threat response

FilletCam AI: A handheld tool for precise fillet color profiling of Atlantic salmon and rainbow trout

Identification of Deadliest Mosquitoes Using Wing Beats Sound Classification on Tiny Embedded System Using Machine Learning and Edge Impulse Platform

Certifications

Online Presence

Get in Touch

AI/ML Engineer
Shipping Production Systems