Arsalan Kamran
ResumeEmail me

Muhammad Arsalan Kamran, AI Engineer

AI engineer with 3+ years building AI SaaS products end to end: agentic systems on LangGraph, DSPy and CrewAI, RAG pipelines, and LLM features on OpenAI, Claude and Gemini. I design the architecture, build the Python backend (FastAPI, Django) and run it on AWS and GCP, taking products from idea to a deployed MVP and on to production.

Email meDownload resume

Based in Peshawar, Pakistan. Working remotely.

Agent run

Which practices are furthest behind on annual wellness visits this quarter?

  1. RouteMetric question. Data API returned no rows, so hand off to the SQL agent.
  2. Select tables
  3. Apply scope
  4. Generate SQL
  5. Guardrails
Illustrative run with sample data, modeled on the agent I built for CareCompass AI.
  • CareCompass AI
  • FreshPass AI
  • BlackGrid
  • Mackinac AI Research
  • TradieMate
  • Locaro.ai
  • ChattyBook
  • Employic
  • MotorCut
  • CRE360.ai
  • Advance Driver Safety

Selected work

Production systems I built or led, and exactly what I was responsible for.

CareCompass AI executive dashboard with savings, wellness-visit and risk metricsCare Compass GPT, the healthcare AI assistant, with recent questions about wellness-visit rates

HealthcareAdvancedACO, current role

CareCompass AI

The problem

ACO executives and care teams need answers from claims and quality data without writing SQL, and clinicians were reviewing faxed home-health plans of care by hand. Every answer has to respect HIPAA and each user's tenant scope.

What I built

  • A LangGraph text-to-SQL agent on BigQuery with LLM table selection, RBAC filter injection and sqlglot validation, used when the primary data API can't answer.
  • An evaluation gate in GitHub Actions that runs a 25-question golden set against every pull request.
  • The 485 document review pipeline: Claude, Gemini and GPT-5 extract and analyze scanned faxes, apply 19 clinical rules, and run as async batch jobs with retries.
  • Service-to-service security with HMAC-signed principals, fail-closed tenant scoping and PHI masking.
  • FastAPI
  • LangGraph
  • BigQuery
  • Next.js
  • Claude
  • Gemini
  • Cloud Run
Visit carecompassai.com
FreshPass AI Tools page with the hairstyle try-on feature

Beauty and bookingIkonic, sole lead engineer for AI and backend

FreshPass AI

The problem

A salon booking platform wanted AI features its customers and salon owners would actually use: seeing a haircut before booking, booking by voice, and asking questions about a business.

What I built

  • A FastAPI service with 42 endpoints delivering 6 AI products, wired into the platform's Laravel backend for billing and audit logs.
  • Hairstyle try-on: face analysis with MediaPipe, a PyTorch face-shape classifier and GPT-4o vision, rendering 3 styles in 4 views each.
  • A real-time voice receptionist on Deepgram that books, checks and cancels appointments through 10 function-calling tools.
  • Multi-tenant RAG assistants on ChromaDB, Redis job queues, retries with backoff and credit debits that roll back on failure.
  • FastAPI
  • GPT-4o
  • Deepgram
  • ChromaDB
  • PyTorch
  • Redis
Visit getfreshpass.com
BlackGrid landing page showing the business score on a phoneBlackGrid AI CFO page with suggested questions and a funding estimate

Case study slides

  • Businesses have the data, not the answers
  • The problem: big questions, scattered answers
  • The idea: start with one business score
  • An AI CFO grounded in the business's own numbers

FintechClentro, lead engineer: full architecture design, AI and backend

BlackGrid

The problem

Small-business owners don't have a CFO, and general chatbots make up numbers. The AI had to answer from the owner's real bank, accounting and email data.

What I built

  • Bootstrapped the Django REST backend and wrote about 54% of its Python code.
  • An AI CFO on the OpenAI Responses API with tool calling, cited web search, streaming and long-term memory, grounded in deterministic cash, valuation and scoring engines.
  • Plaid, QuickBooks, Gmail and Stripe integrations, and iOS and Android apps shipped from the React web app with Capacitor.
  • Django
  • OpenAI
  • Celery
  • Plaid
  • React
  • AWS EC2
Visit getblackgrid.com

Legal researchBoost Panda, contract

Mackinac AI Research

The problem

Policy researchers needed to search state law, cite it correctly and turn their notes into drafts. The prototype's document generator was a mockup.

What I built

  • End-to-end AI document generation: project context from Supabase, GPT-4o drafts streamed over SSE into the editor, and PDF export.
  • A guided multi-turn DSPy agent that collects answers before drafting.
  • Inline citations built only from retrieved sources, and rolling conversation summaries for long sessions.
  • FastAPI
  • DSPy
  • GPT-4o
  • MongoDB
  • React
Visit research.mackinac.org
TradieMate job overview with AI-flagged risks and missing information, client details blurred

ConstructionClentro, lead engineer: full architecture design, AI and backend

TradieMate

The problem

Electrical contractors turn tender emails full of drawings into quotes by hand, across email, spreadsheets and PDF viewers.

What I built

  • A provider-agnostic LLM layer that switches between Gemini, OpenAI and Anthropic.
  • AI quote review and a job-level RAG chat on pgvector, streamed to the React app.
  • Gmail and Drive ingestion that turns a tender email into a verified job and quote.
  • Django
  • pgvector
  • Celery
  • AWS SQS
  • React
Visit the live demo
Locaro landing page with a site viability score

Location intelligenceClentro, sole lead engineer for AI and backend

Locaro.ai

The problem

Business owners in Malaysia wanted to know if an address is a good place to open, using real competitor and population data instead of a guess.

What I built

  • Moved the AI engine from Next.js API routes to Django REST.
  • A two-stage pipeline: Claude Haiku classifies up to 80 nearby competitors, then Claude Sonnet writes a schema-validated viability report.
  • Deployed with Docker on AWS Elastic Beanstalk and ECR.
  • Django
  • Claude
  • Google Places
  • Docker
  • Elastic Beanstalk

Also built

  • ChattyBook reader library
    ChattyBookContinuous ElevenLabs audiobook narration for an AI reading platform.
  • Employic landing page
    EmployicAgentic AI employees with 10+ workplace integrations, including Slack, Google Workspace and Teams.
  • MotorCut landing page
    MotorCutCar photo editing SaaS for dealerships, with AI background removal and image enhancement.
  • CRE360.ai landing page
    CRE360.aiAI market reports for commercial real estate and construction consulting.
  • Advance Driver Safety knowledge hub
    Advance Driver SafetyWebsite and knowledge hub for a fleet driver-safety consultancy.
  • Dev PortalInternal staffing tool with a Gemini agent that reads a project brief and ranks the best-fit engineers.
  • GlobuggDjango backend for a location-based video app, with WebSockets, push and GPT-4o-mini video tagging.
  • VisuAllyObject detection and text-to-speech app for visually impaired users, published in IJIST.

How I work

A repeatable approach for AI products, covering the AI, backend, frontend, CI/CD and infrastructure, from the first conversation to a system running in production.

  1. Discover and scope

    I learn who the users are, what data exists and what success means for the business, then cut the idea down to an MVP that proves value quickly.

    • User flows
    • Data audit
    • MVP scope

  2. Design the whole system

    The AI layer, backend, data model, frontend and cloud setup are planned together up front, so nothing has to be bolted on later.

    • Architecture
    • Model choice
    • Data model

  3. Build AI, backend and frontend

    Agents and RAG with guardrails, clean Python APIs with background jobs, and a web app that streams AI responses to users.

    • LangGraph
    • FastAPI
    • Django
    • Next.js
    • Redis

  4. Test and automate delivery

    AI evaluations, tests and checks run on every pull request, and a merge to main deploys automatically through CI/CD.

    • GitHub Actions
    • LLM evals
    • Docker

  5. Deploy and run in production

    Infrastructure as code on AWS or GCP, with logging, retries and safe rollbacks, then steady improvement based on how people really use it.

    • AWS
    • GCP
    • Terraform

What I focus on most

  • Production-ready from day one

    Every project is built to be deployed, not demoed: containerized services, CI/CD, infrastructure as code and clear failure handling.

  • AI people can trust

    Answers come from the client's own data, with citations, guardrails and validation, so the model doesn't guess when it matters.

  • Security and privacy built in

    Multi-tenant isolation, role-based access and sensitive-data protection, whether the data is patient records, bank transactions or legal documents.

  • Quality you can measure

    Evaluation sets and automated checks show whether a prompt or model change made the product better before users see it.

  • Fast and cost-aware AI

    The right model for each step, small models for simple tasks and swappable providers, so AI stays quick and affordable as usage grows.

  • One accountable owner

    From scoping to launch you work with one engineer who explains trade-offs in plain language and keeps founders and teams in the loop.

Experience

Most recent first. Clentro and Boost Panda were part-time and contract, so they overlap with full-time roles.

  1. Jan 2026 to Present

    Full-Stack AI Engineer

    Advanced Management USA (AdvancedACO), Remote

    Text-to-SQL agents, LLM evaluation and clinical document AI for a HIPAA-compliant healthcare platform.

  2. Dec 2025 to Oct 2026

    Lead AI & Full-Stack Engineer, part-time

    Clentro, Remote

    Led client AI SaaS builds from scoping to release: BlackGrid, TradieMate, Locaro.ai and Globugg.

  3. Dec 2025 to Feb 2026

    AI Backend Engineer, contract

    Boost Panda, Remote

    AI document generation and RAG improvements for a legal research platform.

  4. Sep 2025 to Feb 2026

    AI Engineer

    Ikonic, Islamabad

    Sole lead engineer for FreshPass AI, handling the AI and backend, plus the Dev Portal staffing agent.

  5. Aug 2024 to Sep 2025

    Team Lead, Backend/AI

    CortechSols, Islamabad

    Led a team of 10+ developers across 12+ client projects; RAG systems and Django and FastAPI microservices.

  6. Sep 2023 to Nov 2024

    Machine Learning Engineer

    Ayass BioScience, Remote

    CNN and RNN cancer classification models that reached clinical deployment.

Skills

Tools I have used in shipped work, grouped by where they sit in the stack.

AI and LLM engineering
  • RAG
  • AI agents
  • Text-to-SQL
  • Tool calling
  • LLM evaluation
  • Voice AI
  • Document AI
  • MCP
Frameworks and models
  • LangChain
  • LangGraph
  • DSPy
  • CrewAI
  • OpenAI
  • Claude
  • Gemini
  • Llama
  • Deepgram
  • ElevenLabs
Machine learning
  • PyTorch
  • TensorFlow
  • scikit-learn
  • YOLO
  • ONNX Runtime
  • Stable Diffusion
  • OpenCV
Backend
  • Python
  • FastAPI
  • Django REST
  • Flask
  • Celery
  • Redis
  • WebSockets
  • SSE
Data
  • PostgreSQL
  • pgvector
  • Pinecone
  • ChromaDB
  • MongoDB
  • BigQuery
  • PySpark
  • Databricks
Frontend
  • TypeScript
  • React
  • Next.js
  • Tailwind CSS
  • Capacitor
Cloud and deployment
  • AWS EC2
  • ECS
  • Elastic Beanstalk
  • Lambda
  • Amplify
  • GCP Cloud Run
  • Docker
  • Terraform
  • GitHub Actions
Portrait of Muhammad Arsalan Kamran

About me

I don't stop at the demo. I take a client's AI product from the first idea to real users, then make sure it keeps working once they arrive.

I started in machine learning, training CNN and RNN cancer classification models that reached clinical deployment, and published research on an assistive vision app. Shipping those models taught me that the model is only one part of the product. So I moved into backend engineering, and from there into leading AI and backend work: the architecture, APIs, data pipelines and evaluation that decide whether AI is actually useful.

Today I own the whole path to production. I set up the infrastructure on AWS and GCP (EC2, ECS, Elastic Beanstalk, Lambda, Cloud Run, Docker, Terraform) and the GitHub Actions pipelines that test and deploy every change. After launch I stay with the product, keeping it stable, fast and secure as real users rely on it.

I've led a team of 10+ developers across 12+ client projects and handled scoping and delivery with clients directly, so I can talk through trade-offs with founders as easily as I can debug a production issue.

Your industry isn't on the list? That's fine.

I've shipped AI in healthcare, fintech, legal research, beauty and booking, construction, real estate and location intelligence. These are where I've worked so far, not a limit, and here is why that experience carries over:

  • The hard problems are the same everywhere

    Every AI product needs clean data, grounded answers, security, evaluation and reliable infrastructure. I've solved those across seven industries.

  • I learn the domain from the people in it

    I start with your users and experts to understand the workflow, the vocabulary and the rules, such as HIPAA in healthcare or handling bank data in fintech.

  • Useful work arrives early

    That understanding becomes a scoped plan quickly, so a first working version reaches real users early and improves with their feedback.

Based in
Peshawar, Pakistan (UTC+5), working remotely
Experience
3+ years: ML, then backend, then AI and backend lead
Domains so far
Healthcare, fintech, legal, construction and more, and I adapt to new ones
Education
BSc Computer Systems Engineering, UET Peshawar, CGPA 3.80
Research
VisuAlly, published in IJIST

Have an AI product to build?

I work with founders and product teams in any industry on LLM features, agents and the backends behind them. Tell me the problem, and I'll help you work out how to solve it. Email is the fastest way to reach me.