Senior Software Engineer at BusinessChatAmman, Jordan

Tariq Amarneh

I build AI agents that book appointments, track orders and follow business procedures over WhatsApp, in Arabic and English, and the backend services they run on.

See selected workDownload CVEmail me

WhatsApp conversation

بدي موعد عند دكتور أسنان بكرا الصبح

I need a dentist appointment tomorrow morning

Agent trace

0.00s

0s1s2s3s4s5s

  • guard

    0ms

  • capability_router

  • adaptive_agent

  • └ mentions

  • └ resolve

  • └ live_lookup

  • └ options

  • └ answer

  • final_review

LLM callTool / APIGraph logic

A replay of the kind of agent I build. The spans follow the real graph: a guard, a router, then the agent's own steps. Timings are shortened for the demo. Pick a conversation to watch it run.

Selected work

Five problems I owned from design to production, with the numbers they moved.

BusinessChat, 2026

A booking agent for clinics

guardcapability_routeradaptive loop · driven by the business schemamentionsresolvelive_lookupconfirmcommitanswerfinal_reviewreply
The problem
Patients ask in Arabic dialects, mix doctor names with specialties, and change their mind halfway through. Every clinic's catalog is shaped differently.
What I built
A LangGraph agent driven by each business's schema. Messages pass a guard and a capability router, then an adaptive loop extracts mentions, resolves them against the catalog, checks live availability and confirms before it writes anything.
Result
It runs end to end in production, from the first question to a confirmed appointment.
  • LangGraph
  • LangSmith
  • FastAPI
  • Redis
  • Milvus

BusinessChat, 2026

Making an e-commerce agent 2–5× faster

Before50–80s
After10–40s
0s20s40s60s80s

Response time per customer message, by query type.

The problem
Shoppers waited close to a minute for an answer, long enough to leave the chat.
What I built
Profiled each graph step in LangSmith, ran independent retrieval in parallel, removed redundant LLM round-trips and added fallback models with hard request timeouts.
Result
Responses dropped from 50–80s to 10–40s, depending on the query.
  • LangSmith
  • asyncio
  • Gemini
  • Claude

BusinessChat, 2026

Procedures: workflows written in sentences

Returns.procedure

Ask the customer for @order

/get_order using @order

If it was bought within 14 days,

ask: exchange or refund?

Otherwise, explain the policy

askget_orderjudge ≤ 14 daysyesnoask exchange / refundsay policy
The problem
Businesses wanted their bot to follow their own processes, like returns, escalations and data collection, without filing a request with engineering.
What I built
An editor where admins write steps with @variables and /action chips. Publishing compiles the text into a control-flow graph. At run time, the LLM is limited to four narrow jobs: dispatch, judge, intake and say. A nightly regression suite grades real traces.
Result
Admins build a workflow by writing it out, and the agent runs it the same way every time.
  • Python
  • LangGraph
  • Next.js
  • Kotlin
  • PostgreSQL

PwC, 2024

Semantic search for NEOM

Retrieval accuracy+80%
before
semantic engine
Average search time−65%
before
semantic engine

Measured on NEOM's document library, used by 300,000+ people.

The problem
300,000+ people needed answers buried in a large, growing document library.
What I built
An embedding pipeline and semantic search engine on OpenAI embeddings and ChromaDB that ranks results by context, not just keywords.
Result
Retrieval accuracy improved by 80% and search time dropped by 65%.
  • OpenAI
  • ChromaDB
  • Python
  • Azure

Amazon, 2025

Invoices and 2-hour delivery at Amazon

Each dot is 100,000 items. 15M+ items in Turkey now ship with serial numbers on their invoices.

The problem
Turkish regulations required serial numbers on invoices, and the 2-hour delivery promise in the Gulf needed automatic handling when it was missed.
What I built
Led the serial-number initiative across legal, operations and engineering. Built the auto-cancel and refund workflows for 2-hour delivery in the UAE and Saudi Arabia.
Result
It covered 15M+ items, with better SLA compliance and customer satisfaction scores.
  • Java
  • AWS Lambda
  • EC2

Also shipped

  • Moved the agent runtime to strict mypy, which exposed and fixed 15 latent bugs
  • Root-caused a production memory leak to per-turn gRPC channels
  • Cut LangSmith tracing cost with sampling and payload trimming
  • Generic entity and relation store, so any vertical can sync its catalog
  • Schema translator on Cloud Run Jobs, fed by Pub/Sub events
  • Detection of knowledge-base conflicts, with a review screen
  • IDF-weighted fuzzy name matching for catalog search
  • Nightly regression tiers that grade agent turns from their traces

Career

From forecasting sales at a bank to running AI agents in production. Each bar is a role on the same timeline. Select one to see what I did there.

2020202120222023202420252026

  • Amman

    AI agents for businesses on WhatsApp, from graph to infrastructure.

    • Designed and built the LangGraph agent that took the platform into healthcare. It answers questions about each business's doctors, specialists, services and branches, then books the appointment end to end.
    • Cut e-commerce agent response time from 50–80s to 10–40s.
    • Rebuilt document and FAQ ingestion (loading, chunking, embedding), improving retrieval accuracy by 20%.
    • Built Procedures, which lets businesses write a workflow in plain sentences and compiles it into a graph the agent runs.
    • Python
    • LangGraph
    • LangSmith
    • Kotlin
    • GCP

Stack

Grouped by where it sits in the systems I build, from the channel a customer writes in down to the cloud it runs on.

  1. Channels

    • WhatsApp
    • Web chat
    • REST API

  2. Agents

    • LangGraph
    • LangChain
    • LangSmith
    • RAG
    • Structured output

  3. Services

    • Python · FastAPI
    • Kotlin · Ktor
    • Java · Spring Boot

  4. Data

    • PostgreSQL
    • Redis
    • MongoDB
    • Milvus
    • ChromaDB
    • BigQuery

  5. Cloud

    • GCP Cloud Run
    • Pub/Sub
    • AWS Lambda
    • EC2
    • Azure
    • Docker

  6. Interfaces

    • React
    • Next.js
    • TypeScript
    • Tailwind

Education

BSc, Data Science & Artificial Intelligence

Princess Sumaya University for Technology, 2023

Certificates

Azure AI Fundamentals (AI-900)

Professional Data Analysis, Google Cloud Platform, 2019

Languages

Arabic Native

English Fluent

Portrait of Tariq Amarneh

Get in touch

I'm open to conversations about AI agents, backend systems and engineering roles. Email is the fastest way to reach me.

tariqs.naser@gmail.comDownload CV