Available for consulting

AI & Automation
Engineer building
production LLM
agents & the
backend they
actually run on.

name Abdul Rafay Ahmed
based Islamabad, PK
at Sila Insights · ML & Automation Engineer
since 2020 · 5+ years in production
book a discovery call → see selected work
◆ live demo · ask the agent about my work
rafay-agent · v0.4
idle
// system
rafay-agent v0.4 · ask about my work, stack, or availability
>
scripted responder · for free-form chat, email me
5+
years building in production
Python, GenAI, backend, ML
21
production AI systems
shipped at Sila Insights & Neural Lines
60%
analyst time reduced
via multi-step LLM agent orchestration
100M+
data points processed
across MENA brand intelligence pipelines
40%
LLM inference cost cut
batching, caching, model tiering
350+
hours of training delivered
CI/CD, automation, AI workshops
/01 about

I build the boring infrastructure that makes AI work.

Most AI engineers can ship a notebook. Most backend engineers can ship a service. I do both — which is why my LLM agents don't fall over the moment they leave a demo.

Right now I'm at Sila Insights, an AI consumer-intelligence company in Dubai, where I architect the agent workflows and async data pipelines behind brand-tracking products that ingest 100M+ data points across 20+ Arabic dialects.

Before that, four years across fintech, edtech, and document-intelligence at Neural Lines — early RAG, OCR pipelines, video-CV, and the kind of LangChain plumbing that ships rather than demos.

— Rafay
// what i ship
  • 01
    AI Agent Architecture
    LangChain · LangGraph · LlamaIndex · multi-step reasoning, tool use, evals
  • 02
    RAG Pipelines
    Pinecone · Chroma · Weaviate · FAISS · embeddings · semantic search
  • 03
    Backend Systems
    FastAPI · Django · async Python · Celery · Redis · microservices
  • 04
    ML & NLP
    PyTorch · Transformers · Hugging Face · Arabic dialect models · OCR · CV
  • 05
    Automation
    n8n · Zapier · Airflow · event-driven pipelines · webhook orchestration
  • 06
    Infra & Ops
    Docker · AWS · Azure · Terraform · CI/CD · APM monitoring
/02 selected work

Five projects that actually shipped.

01
year2025 — 2026
clientDataGroomr · Salesforce DQ
roleAI Engineer (contract)
AGENTIC AI · MCP · TOOL ORCHESTRATION
↗ view feature docs

Agentic Data Enrichment Engine

An AI-planned, MCP-native enrichment engine that turns a plain-English prompt into a typed, multi-step execution plan and chains Claude, ZoomInfo, and Apollo across a single Salesforce record.

// problem

Salesforce data rots. Reps inherit half-empty accounts, stale firmographics, and zero buyer-persona context. Manual enrichment is slow, single-vendor lock-in is brittle, and stitching ZoomInfo, Apollo, and an LLM together with hand-written rules turns into an unmaintainable swamp the moment a schema changes.

// approach

Built the agentic core. A planner agent ingests a natural-language prompt with {{Account.Name}}-style tokens, validates them against the source object, then emits a typed, multi-step execution plan: which provider, which tool, which output fields, expected latency, credit cost. Intelligence providers are wired in through their own MCP servers, so tool discovery is dynamic and the planner reasons over live capability lists instead of hand-coded shims. Steps chain output fields forward (firmographics → AI scoring → outreach angle), every plan dry-runs against real records before consuming credits, and execution ships through mass-run, single-record, or real-time create/update triggers, with write-mode guardrails, pre-flight credit checks, and field-level confidence scoring around the whole thing.

Agentic AIMulti-AgentAgentic PlanningMulti-Step ReasoningTool OrchestrationMCPAnthropic ClaudeZoomInfoApolloSalesforceNode.jsVue.js
AI-planned
prompt → execution plan
MCP
native tool discovery
3+
providers chained per run
real-time
create/update triggers
02
year2024 — present
clientSila Insights · Dubai
roleML & Automation Engineer
ARABIC NLP · MULTI-AGENT · REAL-TIME
↗ view product page

Brand Health Tracker

A multi-agent intelligence system processing 100M+ data points to track brand sentiment in real time across MENA markets.

// problem

Analysts at consumer-research firms spend weeks manually compiling brand-perception reports from social media, reviews, and surveys. By the time the report ships, the conversation has moved on.

// approach

Designed a multi-step LLM agent pipeline that ingests data from 6+ social platforms, classifies sentiment across 20+ Arabic dialects, scores brand health on a multi-dimensional rubric, and surfaces alerts via structured API.

LangChainOpenAIPineconeAirflowFastAPIPostgreSQLRedisDocker
100M+
data points / month
60%
analyst time saved
40%
inference cost reduced
20+
dialects supported
03
year2024
clientSila Insights · Dubai
roleML Engineer (lead)
TRANSFORMERS · MENA · FINE-TUNING
↗ view platform page

Arabic NLP Model Suite

Four production transformer models tuned for the messy reality of Arabic dialect classification, emotion, irony, and sentiment.

// problem

Standard Arabic NLP models collapse on dialectal text. Modern Standard Arabic is the language of news, not consumers — and most production models trained on MSA mislabel Egyptian, Gulf, Levantine, and Maghrebi Arabic at humbling rates.

// approach

Curated multi-dialect training corpora, fine-tuned transformer architectures with dialect-aware tokenization, built an evaluation harness that tested across 20+ dialect splits, and deployed via async FastAPI workers behind a Redis queue.

PyTorchHugging FaceTransformersFastAPIRedisDockerHugging Face Hub
20+
dialects classified
4
production models
6M+
training samples
04
year2023
clientNeural Lines · EdTech
rolePython Engineer
EDTECH · LLM AGENT · ADAPTIVE

LLM Exam Simulation System

Adaptive learning agent that generates questions, grades free-form answers, and remembers what each student struggles with.

// problem

Static practice tests don't adapt. Students drill the same easy questions and avoid the hard ones. Tutors don't scale.

// approach

Built an LLM agent with session memory that generates personalized questions on demand, evaluates open-ended answers with chain-of-thought rubrics, tracks performance per concept, and adapts difficulty in real time. GPT API + LangChain + a hand-rolled prompt-eval framework.

GPT-4LangChainFastAPIPostgreSQLRedisReact
1000s
student sessions
real-time
adaptive feedback
0
manual grading
05
year2024
clientSila Insights · Dubai
roleBackend Engineer
BACKEND · ASYNC · IaC

Multi-Platform Social Crawler

Enterprise-scale async crawling across 6 social platforms, feeding the brand-intelligence pipeline.

// problem

Brand intelligence is only as good as its inputs. Sila needed a unified, fault-tolerant ingestion layer pulling from Twitter, Instagram, TikTok, Facebook, Google Reviews, and TripAdvisor — without each platform's quirks leaking into downstream analytics.

// approach

Designed a FastAPI orchestrator backed by Apify actors and Azure Cosmos DB, with Terraform-managed infrastructure, Redis-backed dedup, and platform-specific normalization adapters. Async-first throughout.

FastAPIApifyAzure Cosmos DBTerraformRedisDocker
6
platforms unified
async
end-to-end
IaC
fully managed
/03 project archive

Twenty-one production systems. Plus side projects.

/04 stack

The toolbox, annotated.

/05 experience

Four years, three companies, one direction.

Mar 2024 — Present
ML & Automation Engineer
Sila Insights · Dubai · Remote
Consumer Intelligence · Arabic NLP · Brand Analytics
  • Architected multi-step LLM agent workflows for brand-health tracking — reduced analyst time by ~60%.
  • Built production RAG pipelines (Pinecone, Chroma) over multilingual corpora at 100M+ data-point scale.
  • Engineered prompt frameworks with chain-of-thought reasoning and structured-output guardrails.
  • Cut monthly LLM inference cost ~40% via batching, caching, and model tiering.
  • Documented all agent architectures and integration SOPs for non-technical stakeholders.
May 2022 — Mar 2024
Python Engineer
Neural Lines · Islamabad
Fintech · EdTech · Document Intelligence · CV
  • Shipped LLM-powered exam simulator with personalized learning paths (GPT API + LangChain).
  • Built a Spend Categorization NLP agent with continuous-retraining feedback loop.
  • Trained a real-time person-blurring model: facial recognition + video instance segmentation.
  • Containerized AI agents with Docker; integrated CI/CD for automated testing and rollout.
Feb 2022 — May 2022
Associate Android Developer
O3 Interfaces · Islamabad
Mobile
  • Full lifecycle Kotlin Android development; Play Store beta launches.
  • Optimized cross-device adaptability; reviewed source for production readiness.
Jun 2020 — Aug 2020
Junior Developer (Intern)
CiberSense · Islamabad
Software
/06 teaching · writing

I build and I explain.

// workshops & speaking
How to do AI in 2022
2022
Webinar · GDSC
Automation Testing & CI/CD Pipelines
2023
5-day on-site training · AKSA SDS
AI & ML Leadership
2021–2022
Chapter lead · Google Developers Student Club
// writing — drafts in progress
Why most LLM demos fail in production (and how to fix it)
Coming soon
[Engineering]
Building NLP for 20+ Arabic dialects: the hard parts
Coming soon
[ML]
How we cut LLM inference costs by 40% in 6 weeks
Coming soon
[Cost Eng]
From Android developer to AI engineer
Coming soon
[Career]
/07 recommendations

What people say after we ship.

I worked with Abdul Rafay on the project where he joined as a Python developer and quickly proved himself as a strong and reliable engineer.

He is very good at understanding requirements and picking up new technologies fast. During the project, he had to work with Node.js and Vue.js, and got up to speed with both in a short time without any issues.

Rafay has solid analytical skills — he asks the right questions and thinks through edge cases carefully. He's also comfortable working with AI tools and uses them in a practical way, which really boosts his productivity. On top of that, he's open and easygoing — always willing to help, share knowledge, and support the team.

I really enjoyed working with him and would definitely recommend him.

Roman Sardak
Lead Full Stack Developer · Software Engineer
managed Rafay directly · LinkedIn recommendation, Apr 2026

I hired Rafay at Neural Lines as a Python/ML engineer with limited experience — and watched him grow into someone I trusted with our highest-stakes work. His ability to quickly grasp complex concepts and independently execute has been a standout quality.

For the Red Bull Fortnite tournament, Rafay single-handedly built a system integrating player feeds with webcam streams and tracking eliminations across the bracket. It hit 99% accuracy in production — a testament to his meticulous attention to detail and his ability to ship complex, high-stakes projects on his own.

He pairs that with rare client-communication skills and a knack for turning technical detail into something stakeholders actually understand. A well-rounded engineer with leadership qualities and strong work ethic.

Hassan Jalil
Founder · Neural Lines
hired and managed Rafay directly · May 2022 — Mar 2024
open for new engagements

Have a problem worth
automating? Let's talk.

Best fit: AI-product teams shipping LLM agents to production, founders building MENA-market tooling, or eng leaders who need an AI + backend generalist who actually finishes things.

rafay3515@gmail.com → LinkedIn DM GitHub
// contact card
{
  "name":     "Abdul Rafay Ahmed",
  "title":    "AI & Automation Engineer",
  "based":    "Islamabad, PK · remote",
  "email":    "rafay3515@gmail.com",
  "phone":    "+92 334 5162516",
  "linkedin": "/in/rafay-ah",
  "github":   "@rafay-ah"
}
© 2026 Abdul Rafay Ahmed · built with care, not templates
back to top ↑