AboutCapabilitiesAIExperience WorkContact
Field guide ↗ ScrapeSync ↗ LinkedIn ↗ Résumé ↓
Data Engineer · Lahore, Pakistan · Open to senior roles

I'm Asad Ikram. I build the data systems companies actually rely on.

Chevening Scholar, MSc with Distinction, seven years across six countries. I've built pipelines and platforms for Amazon, Ford, Nike and the UK Government, and I run a data company of my own. Mostly I like the problems other people have given up on.

Chevening Scholar 2024/25 · top 2–3% of 70,000+ applicants · UK FCDO
Scroll
Clients & brands the work reached
Tools I build with
0
Years
0
Freelance projects
0
Scrapers
0
International companies
0
Countries of exposure
How the data flows

From hostile web
to trusted insight

Defended sites in, decision-ready data out — with an AI agent and a quality gate standing guard in the middle.

LIVE FLOW
HOSTILE WEB AKAMAI · CF · DATADOME FETCH LADDER CURL_CFFI FIRST BROWSER LAST S3 RAW PARQUET · CHECKPOINTS DQ GATE SCHEMA CONTRACTS FAIL LOUDLY WAREHOUSE RDL → ODL → ADL REDSHIFT · DBT INSIGHT NLP · LLM · TABLEAU AUTO-HEALER CLAUDE + MCP AGENT ALERTS SLACK · CLOUDWATCH
SWIPE TO EXPLORE →
Track record

Who it was for

Seven years, six countries. Six employers, and work that reached some names you'll recognise.

Where I've worked
Brands & institutions the work reached
About

Two halves of
the same problem

Most engineers pick one. The interesting work lives in the overlap.

01 · The adversarial half

Win the fight before
the page renders

  • JA4 + HTTP/2 fingerprints beaten at the transport layer, before a single byte of HTML
  • curl_cffi first — a patched browser only when the cheap path is genuinely dead
  • Akamai v3 · Cloudflare · DataDome · Kasada · F5 — <1% block rate
HTTP call
Browser40×
the cost gap most teams ignore — and why my proxy bills stay small
02 · The engineering half

Make what you collect
trustworthy

  • Layered warehouses — RDL → ODL → ADL, star and snowflake schemas
  • Schema contracts that fail loudly — never six weeks of silent nulls
  • Monitoring that wakes me up before it embarrasses the client
6 countries · 300+ independent projects · CTO & co-founder of ArtemisAI
What I'm proudest of — uptime, not clever bypasses
70spiders
44domains
0humans in the loop
So I built an AI agent that repairs them on its own
Recognition · UK Foreign, Commonwealth & Development Office
Chevening Scholar 2024/25
The UK government's flagship global scholarship — one fully funded year, awarded on leadership and impact. It's how a scraping engineer from Lahore ended up publishing NLP research in Portsmouth.
70,000+applicants
160countries
2–3%selected
100%funded
Published · The field guide I maintain
Web Scraping Guide
A free, continuously updated reference for the whole discipline — every anti-bot vendor mapped to what actually beats it, and a production Akamai v3 case study with the failed attempts left in. web-scraping-guide.com ↗
73libraries
10anti-bot vendors
9architectures
1honest case study
Education
MSc Data Analytics — Distinction
University of Portsmouth. Dissertation: 279k posts, fine-tuned BERT, macro-F1 0.878. BSc Computer Science, FAST-NUCES.
AI & research

The through-line

Research, production agents and a company, all pointing the same direction. Six stages, and they end somewhere specific.

Capabilities

The full stack

Ten domains. Everything listed here has run in production for a paying client — no coursework, no weekend tutorials.

Experience

Where I've built things

Eight roles, six countries, one throughline. Tap any of them for the numbers.

Selected work

Systems I'm proud of

Six things I'd point to first — three products, an architecture that crawled daily for three years, a published field guide, and the scholarship that reframed the work.

Self-healing systems+
Ford Crawler IQ — Auto-Healer
A spider breaks at 3am. By the time anyone looks, an agent has already read the traceback, pulled the live page, tried four fix strategies, committed the one that worked to a path-locked branch and re-run the crawl. Seventy spiders, forty-four domains, nobody woken up.
Claude SonnetFastAPI MCPDjango DRFECS FargateAurora Serverless
EventBridge weekly ECS Fargate ~70 spiders Django API runs + status Aurora PG serverless v2 spider fails ↓ CloudWatch alert + trigger FastAPI MCP 9 tool endpoints Claude Sonnet agentic loop writes fix ↓ GitHub branch path-locked PR ECS re-run verify + close normal failure heal loop
  • Scope. ~70 Scrapy spiders across 44 car-forum domains, crawled weekly.
  • Guardrails. The agent can only touch spider files on a dedicated branch. It cannot reach pipelines, settings or workflows. Every fix arrives as a PR.
  • Limits. Hard 15-minute timeout at both loop and ECS task level, capped at three heal attempts per spider per week.
  • Escalation ladder. Every fetch walks the cheapest path first: curl_cffi → scrapling → webclaw → plain requests → camoufox → headless Playwright → headful cookie-warming. It stops at the first rung that returns a 200.
  • Cost. Migrated an overprovisioned RDS instance to Aurora Serverless v2, saving roughly $400/month, and 56/56 crawler tests pass on every deploy.
  • Architecture docs ↗
Product+
Where I'm CTO and co-founder. A Meta lakehouse, six fine-tuned NLP models, an agentic analyst chatbot and packaged B2B subscription datasets — with self-healing ops so it keeps running while I sleep.
Custom LLMsRedshiftQuickSight
  • Ingest. GitHub Actions → ECR → ECS Fargate → DynamoDB → S3, pulling 30+ brand pages daily.
  • Warehouse. Layered Redshift (RDL → ODL → ADL) orchestrated by Step Functions with sub-second KPI reporting.
  • Models. Custom LLMs trained on engagement data; predictive trend models lifted client content monetization 40%.
  • Ops. Self-healing EventBridge dead-letter queues, CloudWatch alarms and Twilio alerts — no manual intervention.
  • Access. Multi-tenant Redshift layer with row-level security across 10+ client accounts.
  • artemisai.co.uk ↗
Ops product+
ScrapeYard control plane
One dashboard running a fleet of 86 crawler projects for a parts-retail client — schedule, watch, stop, validate and ship every crawl without touching a terminal. Seven users, zero SSH.
FastAPIAPSchedulerDrive syncCI/CD
  • Registry. 86 manufacturer crawlers with priorities, notes, audit log, soft-delete and restore — the catalogue is the source of truth for what runs when.
  • Drive as ground truth. Google Drive sync walks each crawler's output folder and uses the newest data file's timestamp as the real "last scraped" — no self-reported freshness.
  • Live ops. Colourised streaming logs, run ETAs from windowed completion rates, a real STOP button, and a smoke-test mode capped at 2 models / 25 items for safe dry-runs.
  • Quality gates. On-demand DQ checks — full-file JSON validity, required fields, duplicate detection, diagram coverage — before anything reaches the client.
  • Zero-touch deploys. Push to main, GitHub Actions ships it to the VPS. The client team runs everything through named logins with per-action audit trails.
Ran daily for 3 years+
Dubizzle scraping architecture
500+ spiders on Kubernetes crawling every single day for three years straight across Dubizzle, Bayut, Zameen and OLX — feeding a three-layer warehouse, an Elasticsearch search tier and a competitor-intelligence product built partly from reverse-engineered mobile APIs.
EKSRedshiftElasticsearch
COLLECTION Scrapy fleet 500+ on EKS S3 raw landing Matillion Glue + Athena WAREHOUSE RDL raw ODL ops ADL biz SEARCH Elasticsearch property search +50% query perf INTELLIGENCE Competitor intel MITM + API recon Tableau · Sisense · Retool 15+ stakeholders +60% ETL throughput · promoted twice in two years
  • Fleet. 500+ regional scrapers on EKS with GitHub Actions CI/CD across Dubizzle, Bayut, Zameen and OLX.
  • Warehouse. Redesigned RDL → ODL → ADL for Propforce; throughput up 60%, directly tied to record revenue.
  • Reliability. Introduced unit testing and Scrapy contract validation, cutting production incidents 45%.
  • Recon. Reverse-engineered competitor mobile apps via Android Studio MITM to surface hidden REST and GraphQL APIs.
Published+
The field reference I wished existed — so I wrote it. 73 libraries, 10 anti-bot vendors mapped, 9 production architectures, and a full Akamai v3 case study with the failed attempts left in.
73 librariesAkamai v3JA4
  • Anti-bot map. Ten vendors — Akamai, Cloudflare, DataDome, Kasada, PerimeterX, F5 Shape and more — each with what they check and what actually beats them.
  • The honest case study. A production Akamai v3 bypass documented end to end, including every approach that failed before the one that worked.
  • Deep dives. JA3/JA4 fingerprinting, WASM SIMD detection, mobile API recon, fetch-ladder economics and a priority-ordered tooling cheat sheet.
  • Kept current. Search-indexed, continuously updated as the arms race moves — because a 2024 bypass guide is a history book.
  • web-scraping-guide.com ↗
Chevening exposure+
A year that changed the frame
Chevening isn't just funding — it's a leadership network of 160 countries and direct exposure to how the UK government thinks about data. It turned an engineer into a researcher.
UK FCDO160 countriesTop 2–3%
  • Selection. Chosen from 70,000+ applicants worldwide on leadership and impact — interviews, essays, references, the full gauntlet.
  • The research year. MSc at Portsmouth with Distinction — the dissertation put production scraping skills to academic use: 279k posts, a fine-tuned BERT at macro-F1 0.878.
  • The network. A global cohort of scholars across government, policy and industry — the kind of room a scraping engineer doesn't usually get invited into.
  • The result. UK client work at M+C Saatchi, a UK-registered company of my own, and PhD conversations that started from the dissertation.
Ask AI

Questions? Just ask.

There's an assistant here that has read everything on this page. Free, no login, answers in seconds.

Ask me anything 👋