Senior AI Engineer · Voice AI · RAG · LLM Agents

Ali Armaghan Ahmed

Senior AI Engineer

I design and ship AI products that turn complex workflows into reliable, intelligent experiences — from real-time voice agents and WhatsApp automation to RAG systems, machine-learning pipelines, and AI evaluation frameworks, backed by production-grade backend engineering.

Explore My Projects
voice-agent / live-sessionLIVE
"Hi, it's Alex — large pepperoni on hand-tossed, well done, for delivery please."
{ "intent": "place_order", "order_destination": "delivery", "customer_name": "Alex", "items": [{ "name": "Pepperoni Pizza", "size": "large", "crust": "hand_tossed", "special_instruction": "well done" }], "menu_validated": true, "subtotal": "$16.49", "delivery_fee": "$2.99", "estimated_delivery_time": "35-40 minutes", "order_status": "ready_for_checkout" }
✓ menu-grounded ✓ price-validated ✓ redis session state
7+ YearsData & AI
ProductionLLM Systems
Voice AIConversational Agents
RAGVector Search
ML & DLModel Training
300+Freelance Projects
About

AI systems that work reliably in production — not just in demos.

I am a Senior AI Engineer with 7+ years across machine learning, data science, and AI system development — building practical AI systems that hold up in production. My work spans conversational AI, voice agents, RAG applications, LLM fine-tuning, computer vision, agent evaluation, backend systems, and workflow automation.

I work across the full lifecycle of AI products — from problem design and data pipelines to model integration, API architecture, evaluation, deployment, and monitoring. I care as much about grounding, validation, and failure handling as I do about model capability: an AI product is only useful if it's right when it matters.

That means LLM applications with tool calling and structured outputs, fine-tuning modern models (GPT-4o, LLaMA-3), retrieval systems on Qdrant, real-time voice pipelines at 4,000+ calls a day, satellite-imagery computer vision, FastAPI services with Redis-backed state, Dockerized deployments to GCP and AWS, and evaluation frameworks that catch hallucinations before customers do.

LocationLahore, Pakistan · Remote-friendly
Experience7+ years in Data & AI
CurrentlyML Engineer — CodingKey (Ezra AI)
FocusVoice AI · RAG · LLM Agents
Featured Case Studies

Flagship AI systems, end to end.

Production-oriented builds — real architecture, real constraints, measured outcomes.

Voice AI · Conversational AI · RAG · Restaurant Automation

Turning restaurant calls into accurate, menu-grounded orders — automatically.

A real-time voice assistant for a leading national pizza chain that lets customers place complete food orders naturally over the phone — with menu-aware ordering, pricing validation, payment handling, and human handoff. Not a voice bot: an end-to-end AI ordering system.

The Problem

Meal-time peaks flood the phone lines: long hold times, missed calls, lost orders, and expensive mistakes on complex customizations. Service quality varies by shift, staffing peaks is costly, and after-hours calls go unanswered.

The Solution

A voice-first assistant that answers in real time, understands natural speech, validates delivery addresses, searches the real menu before confirming anything, handles sizes, crusts, toppings and half-and-half requests, suggests relevant add-ons, runs card or pay-at-store workflows, submits the order — and hands off to a human the moment it's needed.

Reliability strategy: the LLM understands language and produces structured order data — it is never trusted to memorize the menu. Item availability, configurations, and prices are retrieved and validated against the live menu before anything is confirmed to the customer.
How a Call Works
  1. Customer calls the restaurant
  2. Voice platform streams speech in real time
  3. Assistant receives transcript + conversation context
  4. Order intent and item details are extracted
  5. RAG menu search retrieves matching items
  6. Order is validated and priced against the live menu
  7. Assistant confirms the order naturally
  8. Payment or pay-at-store workflow completes
  9. Order is submitted — or transferred to a human agent
300+ storesLive in production across two national pizza chains
4–5K calls/dayCustomer calls handled daily across voice and SMS
~60%Of orders completed autonomously, end to end
Architecture Highlights
  • →VAPI for real-time STT/TTS call handling
  • →Qdrant + Whoosh hybrid retrieval (semantic + exact match)
  • →Redis for turn-by-turn memory and cart state
  • →FastAPI orchestration on Docker + Cloud Run
  • →Google Maps address validation · Twilio SMS extension
  • →Flagsmith feature flags for controlled rollout
Technology Stack
PythonFastAPIVAPIAzure OpenAIGoogle GeminiQdrantWhooshSentence TransformersRedisTwilioDockerCloud RunGoogle Maps APIFlagsmith
Download Case Study PDF 🔒 Client name withheld — a leading national pizza chain
Conversational AI · WhatsApp Automation · Lead Intelligence · CRM Automation

From ad click to qualified lead — automatically.

An AI-powered WhatsApp assistant for real estate that turns Facebook and Instagram ad traffic into qualified, CRM-ready leads — engaging every prospect instantly, qualifying naturally through chat, matching properties, and routing high-intent buyers to agents with full context.

The Problem

Click-to-chat leads arrive around the clock, but agents are in calls, meetings, or off-shift. First responses take hours, leads go cold, records arrive incomplete, CRM updates are manual, and agents burn time on prospects who were never going to buy.

The Solution

An intelligent first-touch agent that replies to every inbound WhatsApp lead within seconds, converses naturally instead of pushing forms, detects buy-vs-rent and personal-vs-investment intent, captures budget, location, and property requirements, matches live listings via dynamically generated SQL, books calls or visits, and hands interested leads to sales with the complete transcript.

Conversation Design
  1. Warm welcome + initial intent detection
  2. Buy or rent qualification
  3. Purpose: personal use, investment, or business
  4. Budget discovery
  5. Location & property preference collection
  6. Property matching from the live database
  7. Top matching properties presented
  8. Call or office-visit scheduling
  9. Lead classification + sales-agent handoff
Interested

High-intent lead: clear requirement, defined budget, preferred location, and willingness to schedule a call or visit.

Follow-Up Needed

Engaged prospect with real interest but no immediate booking or commitment.

Not Interested

Low-priority browsing or pricing-only contact, retained for future nurture and remarketing campaigns.

< 90 secFirst response time — down from 3–6 hours
3×More fully qualified leads per day
68%Less agent time spent on cold leads
24/7Coverage, including weekends and holidays
100%Complete CRM entries at creation
+41%Scheduled sales calls, month over month
Human Handoff — what agents receive

Lead name · phone number · preferred contact time · budget · location · property requirements · properties shown · AI-generated classification · full WhatsApp conversation transcript.

Technology Stack
PythonFastAPIWhatsApp Business APIMeta Click-to-ChatOpenAI / LLM APIMySQLDynamic SQL GeneratorGoogle Sheets APICRM APIWebhooksAgent RoutingLead Classification
Download Case Study PDF 🔒 Confidential client project — presented as a case study
More Projects

Selected engineering work.

AI Evaluation · Agent Reliability · LLM Testing

AI Agent Evaluation & Reliability Framework

A framework for testing AI agents against realistic scenarios, tool failures, hallucination patterns, instruction-following requirements, and multi-step workflows.

  • Tool-call trace evaluation & hallucination detection
  • Failure simulation and error injection via middleware
  • Rubric-based grading + database-state validation
PythonLLM APIsTool CallingMiddlewarePostgreSQLDocker
Speech AI · Machine Learning · Fine-Tuning

Arabic Whisper Speech Recognition Fine-Tuning

Fine-tuning and evaluating Whisper-based speech-recognition models for Arabic and Quranic-recitation use cases.

  • Dataset prep, audio normalization & GPU training
  • Word Error Rate evaluation across model variants
  • Faster-Whisper inference & streaming exploration
PythonPyTorchHugging FaceWhisperFaster-WhisperCUDARunpod
LLM Automation · RAG · Workflow Automation

AI-Powered Email Order Automation

An automated email-processing system that extracts customer orders, identifies products, validates stock, updates inventory, and generates grounded customer responses.

  • Email classification + LLM structured extraction
  • Product matching via vector search
  • Inventory validation & grounded auto-responses
PythonFastAPIQdrantSentence TransformersLLM APIsPostgreSQL
Skills & Technical Expertise

Depth where AI products actually break.

LLM & Agents

  • LLM application design
  • AI agents & orchestration
  • Tool calling
  • Structured outputs
  • Prompt & context engineering
  • OpenAI · Azure OpenAI · Gemini

RAG & Retrieval

  • RAG architecture
  • Qdrant · Chroma
  • Whoosh keyword search
  • Hybrid semantic + exact match
  • Sentence-transformer embeddings
  • Grounded answer generation

Voice & Speech

  • Real-time voice agents
  • VAPI · LiveKit
  • STT / TTS pipelines
  • Whisper fine-tuning
  • Faster-Whisper inference
  • Telephony (Twilio)

Machine Learning

  • Deep learning (PyTorch)
  • Model fine-tuning
  • Hugging Face Transformers
  • GPU training workflows
  • WER / model evaluation
  • Experimentation & analytics

Backend & APIs

  • FastAPI service design
  • Node.js
  • Webhooks & integrations
  • WhatsApp Business API
  • REST API architecture
  • Auth & token services

Data & Storage

  • Redis (state & memory)
  • MySQL · PostgreSQL · SQLite
  • Dynamic SQL generation
  • Google Sheets API
  • Data pipelines
  • CRM integrations

DevOps & Cloud

  • Docker
  • Google Cloud Run
  • CI-friendly deployments
  • Runpod GPU infra
  • Feature flags (Flagsmith)
  • Monitoring & logging

Evaluation & Reliability

  • Agent evaluation frameworks
  • Hallucination detection
  • Failure simulation
  • Error injection & recovery
  • Rubric-based grading
  • DB-state validation
Experience

Seven years, from data analytics to production AI.

Oct 2023 — Present · Remote

Machine Learning Engineer — Voice & Conversational AI

CodingKey (Ezra AI)

Built and optimized LLM-powered voice and chat ordering agents (Azure OpenAI, Google Gemini) for two national pizza chains — live across 300+ stores and handling 4,000–5,000 customer calls daily. Designed the RAG menu-intelligence layer (Qdrant, Whoosh, sentence transformers), FastAPI microservices with Redis-backed conversation and order state, Twilio/VAPI telephony and payment integrations, and Docker + Cloud Run deployments. Led prompt engineering, evaluation, and supervised fine-tuning that brought the AI to ~60% fully autonomous order completion.

Aug 2023 — Aug 2024 · Lahore

Data Scientist — Computer Vision & Multimodal AI

TenX

Led an end-to-end computer-vision pipeline on Sentinel-2 satellite imagery to forecast Pakistan's national cotton production — 86.88% forecast accuracy, delivered ~3 months before harvest — directly informing client procurement, inventory, and import/export planning.

Jan 2021 — May 2023 · Lahore

Data Analyst — Healthcare Analytics

Emerald Labs

Delivered end-to-end analytics for an 8+ clinic US healthcare network: Power BI and Google Data Studio dashboards on optimized SQL, insights on capacity, appointment trends and no-shows, plus master-data governance and access control for confidential patient data.

May 2018 — Dec 2020 · Lahore

Data Analyst — MIS & Forecasting

Provincial Health Department, Punjab

Managed patient and medicine-consumption analytics across 175+ clinics; built 6–12 month medicine-forecasting models that reduced shortages, overstocking, and unnecessary spending — and architected the department's Management Information System end to end.

AI Voice Assistant

Don't read about my work. Talk to it.

This portfolio ships with its own real-time voice agent — built on the same architecture as my client systems. Ask it about my projects, stack decisions, availability, or how I'd approach your problem, and it answers in real time.

LiveKit real-time audioLLM + RAG over this portfolioFastAPI token service
Tap to start a voice session
Contact

Have an AI product to build?

I work with founders, product teams, and engineering leaders on voice agents, RAG systems, LLM automation, and AI reliability. Tell me what you're building — I'll tell you how I'd ship it.

Email Me LinkedIn WhatsApp GitHub Talk to My AI Assistant