Hamza Aziz

AI Product Engineer at VoicePlug Inc., with experience across Computer Vision, Generative AI, and Conversational/Voice AI.

Open to AI engineering roles in the UAE and India. Based in India, ready to relocate to Dubai or Abu Dhabi with visa sponsorship.

Hiring outside the Gulf? Get my AI Engineer resume (PDF)

View CV Download CV

Scroll Down

About Me.

AI Engineer with a passion for building end-to-end AI solutions using Computer Vision, Generative AI, and Conversational/Voice AI. With 3 years of hands-on experience, I've shipped real-time facial recognition systems, voice ERP assistants, and a production drive-thru voice ordering agent — delivering scalable and production-ready AI products.

My Skills Are: Python, Computer Vision, Generative AI, Conversational/Voice AI (LangChain, LangGraph, Rasa, Gemini), Agentic AI (ReAct, Function Calling), Deep Learning (PyTorch, TensorFlow, TensorRT), Databases (Milvus, MongoDB, PostgreSQL, Redis), SQL, Docker, REST APIs, Git & GitHub, Real-time Inference, Edge AI (Jetson Orin Nano, Raspberry Pi)

Contact Me

Skills.

Programming and AI/ML

Python

Expert

Rasa

Intermediate

LLM (Gemini)

Intermediate

Hugging Face

Intermediate

LangChain

Intermediate

TensorFlow

Intermediate

PyTorch

Intermediate

OpenCV

Intermediate

TensorRT

Intermediate

Web Development

HTML

Intermediate

CSS

Intermediate

JavaScript

Beginner

Flask

Intermediate

Fast API

Intermediate

Deployment and Containerization

Docker

Intermediate

DeepStream SDK

Intermediate

NVIDIA Jetson

Intermediate

Raspberry

Beginner

Database and Backend Technologies

Milvus: Vector Database

Intermediate

Confluent Kafka

Intermediate

PostgreSQL

Intermediate

Redis + Python

Intermediate

Neo4j + Python

Beginner

Qualification.

Education

Master of Technology - Artificial Intelligence and Machine Learning

Birla Institute of Technology and Science (BITS) 2025 - 2027

Bachelor of Technology - Computer Science & Engineering

ABES Institute of Technology, AKTU 2019 - 2023

Senior School Certificate Examination

Green Field Public School, CBSE 2017 - 2019

Work Experience

AI Product Engineer

VoicePlug Inc. September 2025 - Present

Assistant Engineer - AI

Global InfoVentures Pvt. Ltd. August 2023 - August 2025

AI Intern

Global InfoVentures Pvt. Ltd. February 2023 - July 2023

Machine Learning Intern

Feynn Labs September 2023 - November 2023

Projects.

VoicePlug Inc.

DriveThru Voice Ordering Modernization

A real-time voice agent that takes DriveThru orders and files them into the POS, across multiple QSR brands. On FastAPI + Socket.IO, a speech layer streams the customer's words in, Gemini 2.5 Flash-Lite extracts menu items, an order agent replies from a cached menu-aware prompt, and a second model turns each spoken confirmation into Redis cart operations. The hard part is making a non-deterministic model safe to bill on: required-choice and per-item limit enforcement, a final cart check before submission, and staff handoff after repeated failures — deterministic code, validated against replays of real calls. MongoDB records, Jinja2 per-brand prompts, four POS integrations.

Read case study

Final-Year Project | ABESIT

Mr. Healthcare - The Medical ChatBot

A medical question-answering chatbot over a Neo4j knowledge graph of 658 diseases that says plainly when it cannot answer. A Vue 3 page talks to a Flask API, which finds disease and symptom names with an Aho-Corasick automaton and answers only through fixed, parameterized Cypher queries, with follow-ups (“what causes it?”), clarifying questions and suggested next questions. A language model (gpt-oss through Ollama) may read each question to handle paraphrases and negation, but it only chooses among the names the rules found and never writes an answer at question time; if it is slow or unsure, the rules answer. On a fresh, frozen set of 120 scenarios it lifted complete understanding of paraphrases and follow-ups from 54.7% to 70.9% (+16.3 points) with no new wrong-subject answers, but missed its reliability (98.1% usable replies against 99%) and latency (p95 1.56 s against 1 s) targets. Path-filtered GitHub Actions CI with a knowledge-graph gate, automatic releases and Docker images on GHCR. Built with Mahak Agrawal. Live demo; not medical advice.

View Live demo

Personal Project

HalalLens - The Shariah Compliance Screener

A full-stack Shariah-compliance screener for BSE-listed Indian equities: a FastAPI backend and a Next.js 16 front end over PostgreSQL 17. Every verdict, from the API or a scheduled job, is published by one function under a Postgres advisory lock; a missing input gives insufficient data rather than a guess, an open data-quality issue gives needs-review, and a compliant verdict needs an admin-supplied prohibited-income base. The “Explain Why” narrator runs on deterministic rules by default, citing the filing date and figures, with an optional vLLM / OpenAI-compatible mode. Subscribers get compliance-flip alert emails with at-least-once delivery, and a Razorpay billing ledger, on test keys only, handles plans, refunds, a reconciler for missed webhooks and credited Lifetime upgrades. Releases publish Docker images, deployed to an Oracle Cloud VM behind a Cloudflare Tunnel, with nightly backups and uptime checks; 1,092 backend and 47 UI tests in CI, plus opt-in real-Postgres locking tests. Production holds no BSE filings yet, while a permitted data source is pending, and the screening thresholds are not scholar-validated.

Source code halallens.in

Open Source

Codex–Claude Council — Two AIs, One Answer

An open-source plugin and MCP server that puts one question to both ChatGPT (via the Codex CLI) and Claude (via Claude Code). They answer independently, then reply to each other round after round until they agree — one model writes the final answer while the other checks it. Built on Node.js 20+ and speaking MCP over JSON-RPC/stdio, it runs locally on your own plans by default; either side can take an API key instead, or skip its CLI and run through the OpenAI or Anthropic API. It works across Claude Code, Codex, Claude Desktop and the ChatGPT app. On a 24-question benchmark of deliberately tricky programs, run from fresh sessions with no web search, it scored 24/24, matching the best single model while catching the other's one miss.

View npm

Global InfoVentures Pvt. Ltd. | Team project

Conversational AI for College ERP

A voice- and text-driven assistant that answers questions against live college ERP data across multiple client institutes. Audio is transcribed server-side through a FastAPI gateway, then a two-stage Rasa router picks the ERP module — asking a clarifying question and re-classifying when confidence is low, so an ambiguous query never reaches the LLM. Schema grounding is the hard part: each institute's hundreds of field mappings are embedded into Milvus and retrieved alongside similar few-shot examples, so Gemini 2.5 Flash emits constrained SQL fragments into a fixed payload contract instead of free-form SQL. The ERP executes it and a second model summarises the rows.

Read case study

Global InfoVentures Pvt. Ltd.

PTZ Classroom Attendance — One Camera per Room

Classroom attendance from a single PTZ camera per room instead of a wall of fixed cameras. The system drives one ONVIF camera through a timed preset sweep — dedicated faculty captures at lecture start and end, then a 23-preset trace across the seats — with zoom and focus settling budgeted into every hop. Frames go to a TensorRT-backed InsightFace-REST service for detection and embedding, and identities come from a top-1 Milvus search over 512-d vectors. A separate YOLOv9 service counts heads independently; the two are reconciled per lecture and synced to the institute ERP. Four Dockerised services on a MIG-partitioned GPU, with PostgreSQL attendance logs.

Read case study

Global InfoVentures Pvt. Ltd.

CCTV Face Recognition Attendance — Multi-Stream

The multi-stream sibling to the PTZ system. A producer process decodes 20+ live RTSP camera feeds — ROI crop and lighting check per frame — into a shared in-memory queue; a consumer batches 32 frames per request to InsightFace-REST, running as two TensorRT replicas on separate MIG slices behind nginx, where SCRFD (scrfd_10g_gnkps) detects and w600k_r50 embeds. Identities resolve by cosine search over a GPU-indexed Milvus collection. The gallery improves itself: high-confidence crops land in a staging collection that a nightly job folds into the index. Results persist to PostgreSQL, sync to the institute ERP, and stream back to on-site displays over Kafka.

Read case study

Personal Project

Clinic Voice Receptionist — Multilingual Voice Booking

An inbound voice agent for dental clinics, built against a real published brief, that books appointments in English, Hindi, Tamil, Kannada and Malayalam and follows up over WhatsApp. On FastAPI, a local Whisper model on the GPU transcribes each turn, an LLM slot extractor (LangChain + Ollama, with a rule-based fallback) pulls out the booking details, and Piper and MMS-TTS speak the reply — 0.8–1.0 s per turn on an RTX 4050. The rule that matters: details it heard with low confidence are read back to the caller before anything is booked. On its own 11-utterance test set with simulated transcription errors up to 50%, accuracy falls to 74.4% but no wrong detail goes unflagged. Telephony (Bolna), WhatsApp (AiSensy) and n8n routing are built against their documented payloads.

View

Personal Project

AgencyOps Orchestrator — Agentic Workflows, Human Approval

Agentic workflow automation for an eCommerce marketing agency, built against a real published brief: it drafts weekly client performance reports and ad-creative variants. LangGraph state graphs with bounded revise loops do the drafting, and nothing reaches a client or an external tool until a person approves it in a FastAPI review console — workflows only propose external writes, and one dispatch path runs them after approval. All arithmetic is deterministic Python; the LLM (Ollama or Gemini, with an offline fallback) only writes the prose, and a verify gate rejects any narrative that contradicts the numbers, such as “CPA fell” when it rose. Meta Ads, Harvest, Slack and Trello connectors; the demo and tests run offline with no credentials.

View

Personal Project

Dubai Smart Irrigation Advisor — Ranked by Cost, Not Error

Built against a real published brief, it decides when to irrigate in Dubai's desert climate, and ranks six predictors by what they cost to run in AED rather than by prediction error. The physics core is FAO-56 Penman–Monteith ET0 with a root-zone soil water balance; it reproduces FAO-56's worked examples exactly and agrees with published Dubai climate normals to 2.3% over 30 years of NASA POWER data. Against it run scikit-learn, XGBoost and PyTorch LSTM/GRU models, each scored by what its decisions cost in water and AED over a simulated 120-day season. The best learned model comes within 5% of the physics baseline, which also drives the simulator; acting on the raw soil sensor alone costs 6.7 times as much. A Streamlit dashboard shows the decisions, with optional plain-language explanations from a local LLM whose every number is checked against the facts.

View

Personal Project

Career Growth AI for Kids — Built for Children's Data

A web app, built against a real published brief, where parents build profiles for their children — age band, grade, interests and strengths — and get AI-generated career-exploration ideas. Children never log in, and parents own all the data. FastAPI with async SQLAlchemy over PostgreSQL, a React 19 + TypeScript front end, and an Ollama-served model. The care went into handling children's data: a child's name and the parent's notes are never sent to the AI (tests enforce it), only the birth year is stored, and a parent asking for another parent's child gets a 404. Auth uses short-lived JWTs in httpOnly cookies and hashed, rotating refresh tokens — reuse one and every session is revoked — with per-user rate limits on login and AI generation.

View

Internship | Global InfoVentures Pvt. Ltd.

PPE Violation Detection

A browser-based PPE monitor: point it at an uploaded clip or a live camera URL and it streams the raw and annotated feeds back side by side over MJPEG. A custom-trained YOLOv7 model classifies every person in frame as compliant or violating at a 0.75 confidence gate, and a temporal rule waits for a run of unsafe frames before firing — so one stray detection doesn't page anyone. When it does fire, the Gmail API emails the offending frame to a configured recipient as evidence, then holds a cooldown before the next alert. Flask, OpenCV and PyTorch on CPU, with per-run detection reports exportable from the UI as a timestamped log.

Contact Me.

I will read all emails. Send me any message you want, and I'll get back to you as soon as possible.

I need your Name and Email Address, but you won't receive anything other than your reply.

Send Me A Message

Does not send emails

Write me on my social networks