HELLO I'M
HAMZA AZIZ

Real-time voice agent that takes DriveThru orders and files them into the POS across multiple QSR brands. On FastAPI + Socket.IO, Gemini 2.5 Flash-Lite extracts menu items, an order agent replies from a cached menu-aware prompt, and a second model turns spoken confirmations into Redis cart operations. Deterministic guardrails make it safe to bill on: required-choice and per-item limits, a final cart check, and staff handoff after repeated failures, validated against replays of real calls. Live at 2 restaurants with a 2.63 s median reply over 17,884 turns; MongoDB records, Jinja2 per-brand prompts, 4 POS integrations.

Built and shipped a Shariah compliance screener for halal (Islamic) stock screening of BSE-listed Indian equities: five AAOIFI-inspired ratios (debt, cash, non-permissible income, receivables, liquid assets), a three-state verdict (compliant / non-compliant / insufficient data) and rule-based “Explain Why” narration. Every verdict publishes through one locked transaction, so a known data-quality issue never surfaces as compliant or non-compliant. FastAPI + Next.js 16 + PostgreSQL 17 on an Oracle Cloud arm64 VM behind a Cloudflare Tunnel; CI runs 1092 backend and 47 UI tests and gates multi-arch Docker images with smoke tests and a Trivy critical-CVE scan.

Open-source MCP server and plugin (Claude Code, Codex, Claude Desktop, ChatGPT app) that has ChatGPT (via the Codex CLI) and Claude (via Claude Code) answer a question independently, then debate in lock-step rounds until both agree; one writes the final answer, the other verifies it. Each side runs through its CLI (Codex, Claude Code) on the user's own plan by default or on that side's API key, or, with no CLI installed, through the OpenAI Responses or Anthropic Messages API with read-only project tools and web search. Node.js; 8 MCP tools plus prompts and resources; remote sessions over SSH; Streamable HTTP over a Cloudflare tunnel; multi-arch Docker image on GHCR and Docker Hub; 118 automated tests on Windows/macOS/Linux CI; tokenless OIDC releases to npm and the MCP Registry.

A medical question-answering chatbot whose every answer comes from a Neo4j knowledge graph built from a disease dataset of 658 diseases (symptoms, causes, prevention, treatments, complications, descriptions); questions it cannot answer get an explicit “I don't know”. A language model may help read a question but never writes the answer. Vue 3 + Vite frontend and Flask API (gunicorn), shipped as Docker images on GHCR and Docker Hub behind nginx; GitHub Actions CI with an evaluation gate before every release, and self-updating deploys with automatic rollback on an Oracle Cloud VM behind a Cloudflare Tunnel. Not medical advice.

Worked in the team that built a voice and text query assistant that answers natural-language questions over a multi-institute college ERP. A FastAPI backend transcribes speech via Google Speech-to-Text, routes queries to student, fee, employee or library modules with two Rasa DIET classifiers, and asks a clarifying question when confidence is low. Schema fields (100+ field-level mappings per module) and few-shot examples are retrieved from Milvus using MiniLM embeddings, improving SQL precision by 30%+, with vector-based field/value disambiguation adding 40%+. Gemini emits the query fields and conditions, the application builds the ERP payload rather than full SQL, and results are summarized into gTTS speech; near-100% intent classification accuracy.

Built real-time attendance on two camera setups sharing one recognition stack. CCTV: 20+ live streams on an NVIDIA DGX A100 using MIG partitions, with >99.99% accuracy against the attendance register in controlled UAT at the main and library gates. RTSP streams are batched into InsightFace-REST, which serves TensorRT SCRFD (scrfd_10g_gnkps) detection and w600k_r50 recognition as two replicas on separate MIG slices behind nginx. Dual Milvus collections pair the enrolled gallery with a CCTV-captured staging gallery to counter domain shift; a collector keeps each person's best faces, ranked by pupil distance and match score, and a nightly job refreshes the vector DB, raising real-world accuracy from 85% to 99%. PTZ classroom: one PTZ camera per classroom sweeps up to 23 ONVIF presets over RTSP, with dedicated faculty captures at lecture start and end; a YOLOv9 service estimates headcount, FastAPI services write lecture-wise attendance into year-partitioned PostgreSQL tables, and an APScheduler job syncs to the institute ERP every 5 minutes across 8 lecture slots per day. Also: Docker, Kafka feedback, DVC-versioned datasets.