Junior AI Engineer @ Brilliant Infotech

AI Engineer building production-grade conversational AI, RAG pipelines, and multi-agent systems.

M.S. in Data Science from NJIT. Currently at Brilliant Infotech, building an AI-powered talent-matching platform on the Claude API — on top of a data-engineering foundation from Verizon (2B+ telemetry records) and JerseyStem. Independently, I build things like CareFlow, a security-first voice AI receptionist — live demo below.

View Projects Get in touch
Harrison, NJ M.S. Data Science, NJIT — 2025 Building CareFlow, an AI voice receptionist

LLM & AI Engineering

LLM APIs (Claude, GPT) RAG Multi-Agent Systems LangChain HuggingFace

ML / Modeling

PyTorch TensorFlow Scikit-learn XGBoost Keras

Data Engineering

PySpark Azure ADF / ADLS DuckDB MySQL / BigQuery

BI & Cloud

Looker Studio Power BI / Tableau AWS · GCP · Azure
01 · Projects

Systems I've designed and shipped

Led by CareFlow, a production-shaped voice AI system — followed by a multi-agent LLM research platform and RAG-driven analytics. Ordered by how directly they reflect AI engineering work.

Caller voice LiveKit + Deepgram GPT-4o-mini FastAPI auth-gated Booking / Calendar Supabase pgvector RAG · clinic FAQ OWASP LLM Top 10 red-team tested
FLAGSHIP

CareFlow AI — Voice AI Medical Receptionist

A security-first voice AI receptionist for medical clinics — booking, canceling, and rescheduling appointments through natural conversation, with the LLM treated as an untrusted client at every step.

  • Architected a security-first design where all identity verification and authorization run server-side in FastAPI — no booking, cancellation, or reschedule can execute without a backend-validated session.
  • Built a real-time voice pipeline in LiveKit (Deepgram STT, GPT-4o-mini, OpenAI TTS), validated by 29 automated tests covering identity, tool-calling, and booking flows.
  • Designed a RAG-based FAQ system on Supabase pgvector for clinic policy Q&A, with a similarity threshold tuned to prevent hallucinated answers on out-of-scope questions.
  • Ran an OWASP LLM Top 10 red-team evaluation against the tool layer — prompt-injection and authorization-bypass vectors — with all findings remediated before deployment.
PythonFastAPILiveKitSupabase / pgvectorDeepgramOpenAI
Live Demo
Claude Haiku GPT-4o-mini Mistral Fast reflex loop Planning + RAG Memory Redis Pub/Sub FastAPI Telemetry — 5,000+ action-level events Bronze Silver Gold Databricks
02

Minecraft Multi-Agent LLM Research System

A PIANO-inspired multi-agent architecture comparing how different LLMs behave, cooperate, and specialize during autonomous, open-ended gameplay.

  • Designed a dual-loop architecture — a fast reflex loop for immediate survival reactions and a slower LLM planning loop — comparing Claude Haiku, GPT-4o-mini, and Mistral across 6+ hours of autonomous play; agents developed distinct explorer vs. homekeeper strategies with no explicit role assignment.
  • Built RAG-based episodic memory (HuggingFace Sentence Transformers), Redis pub/sub inter-agent coordination, and an async FastAPI inference server — agent performance improved 78% session-over-session (9→21 unique items).
  • Captured 5,000+ action-level telemetry records through a Bronze/Silver/Gold pipeline in Azure ADLS Gen2, then ran cross-model behavioral analysis in Azure Databricks/PySpark comparing survival, exploration entropy, and resource efficiency.
PythonNode.jsFastAPIRedisHuggingFaceAzure Databricks
GitHub
CMS Medicare Claims 20,867 records Feature Engineering XGBoost Model 0.908 AUC 85+ age: 57.9% readmit Mental health: 57.6% ChromaDB + Claude API RAG Chat Interface Streamlit
03

ReadmitIQ — Hospital Readmission Analytics Platform

An end-to-end ML + RAG platform predicting 30-day hospital readmission risk from synthetic CMS Medicare claims, with a natural-language query layer on top.

  • Built an ML pipeline on CMS synthetic Medicare data (58,066 raw claims → 20,867 deduplicated, generated via the Synthea simulation framework — no real patient data); XGBoost reached 0.908 AUC vs. a 0.772 logistic-regression baseline.
  • Engineered 15+ clinical features (length of stay, admission type, diagnosis category, demographics) and flagged the 85+ age group (57.9% readmit rate) and mental health diagnoses (57.6%) as highest-risk segments.
  • Chose DuckDB over Postgres for in-process speed, and a ChromaDB-based RAG layer over live SQL generation specifically to eliminate hallucination — enabling a plain-English chat interface (Claude via LangChain) over the analytics.
PythonXGBoostDuckDBChromaDBLangChainStreamlit
GitHub
Match Scraper 30+ matches MySQL 4-table schema SQL KPI Views ACS · KAST · ADR Looker Studio
04

VCT Pacific Data Pipeline & Competitive Analytics

An automated scraper-to-dashboard pipeline turning raw esports match data from vlr.gg into a coaching tool used to prep for real matches.

  • Built an automated ETL pipeline scraping match schedules, results, and player statistics for 30+ VCT Pacific matches into a normalized MySQL schema, with deduplication and a hybrid daily-incremental / weekly-full refresh strategy.
  • Engineered SQL views computing advanced KPIs (impact score, K/D ratio, kill contribution, rolling averages) across metrics including ACS, KAST, ADR, and first-blood efficiency.
  • Built a Looker Studio coaching dashboard for ATK/DEF win rates, player heatmaps, and opponent compositions — validated by a professional VCT analyst.
PythonBeautifulSoupMySQLLooker Studio
GitHub
Earlier work — data analysis & ML foundations
Jan – Mar 2025

NBA Team Composition Analysis & Optimization

Scraped and cleaned NBA player data (BeautifulSoup, Selenium) and ran EDA in Power BI / Plotly to visualize salary cap usage, player efficiency, and undervalued trade targets.

PythonPower BIPlotly
Jan – Mar 2024

Image Generation with Conditional GAN

Developed and tuned a Conditional GAN on CIFAR-10 to generate high-fidelity, class-conditioned images; evaluated with inception score and visual inspection.

PythonPyTorchGANs
Oct – Dec 2023

Big Data Classification Analysis: CIFAR-10

Comparative analysis of CNNs, linear regression, and single-layer networks for image classification, benchmarked with cross-validation, accuracy, and F1-score.

PythonCNNsML
02 · Experience

From data foundations to AI engineering in production

Currently building AI product features professionally at Brilliant Infotech, on top of the data-engineering foundation from JerseyStem and Verizon.

Jul 2026 — Present
Current

Brilliant Infotech

Junior AI Engineer
  • Supporting development of an AI-powered talent-matching platform built on Next.js, Supabase/Postgres, Prisma, and the Claude API — contributing across data sync, scoring, and pipeline reliability.
  • Debug and resolve defects in candidate data pipelines, including third-party ATS integration syncing, deduplication logic, and eligibility-extraction rules, with regression tests in TypeScript.
  • Collaborate with senior engineers on data modeling and system integrations supporting recruiting and account-management workflows.
Aug 2025 — Jul 2026

JerseyStem

Data Analyst
  • Built MySQL-based data pipelines replacing fragmented spreadsheets across 3 departments with a single verified data source, reducing manual cross-checking.
  • Streamlined Salesforce and GAT+ activity data into centralized reports, letting contributors prioritize outreach in real time and eliminating recurring manual data requests.
  • Partnered with Marketing, Operations, and leadership to design and deploy 4 Looker Studio dashboards covering pipeline activity, team performance, and school outreach metrics.
Feb 2025 — May 2025

Verizon

Data Analyst
  • Engineered high-performance ETL/ELT pipelines in Python and PySpark processing 2B+ telemetry records, optimizing ingestion, transformation, and compression while maintaining data integrity.
  • Designed automated ingestion pipelines for multi-source network data with validation frameworks and quality checks, storing cleaned data in Parquet on university HPC for next-day analysis.
  • Conducted large-scale EDA to identify usage trends and performance anomalies — findings that directly informed infrastructure planning presented to cross-functional engineering teams.
03 · About

I build the AI layer on top of a solid data foundation.

I'm an AI Engineer with an M.S. in Data Science from NJIT. I'm currently at Brilliant Infotech, building an AI-powered talent-matching platform on the Claude API, Next.js, and Supabase. Before that, I built data pipelines at JerseyStem and processed 2B+ telemetry records at Verizon.

My independent project work — CareFlow's security-first voice AI, a multi-agent LLM research system, RAG-driven analytics — reflects where I want to keep building: production AI systems grounded in rigorous data engineering, not detached from it. That progression, not a clean break from data analytics, is the throughline of my work.

2B+
Telemetry records processed (Verizon)
0.908
AUC — ReadmitIQ readmission model
78%
Session-over-session agent improvement
29
Automated tests validating CareFlow's voice AI flows
New Jersey Institute of Technology
Master of Science, Data Science
Sep 2023 — May 2025 · Newark, NJ
04 · Contact

Let's talk

Always happy to connect about AI engineering, LLM systems, or interesting problems to work on. The fastest way to reach me is email or LinkedIn.

Email harshadhatteusa23@gmail.com LinkedIn linkedin.com/in/harshad-hatte GitHub github.com/Harshad30