I'm a Product Designer, not a cloud engineer — but I have several close friends in Lahore who are DevOps and Cloud Engineers actively job-hunting, and I kept watching them burn hours every day scrolling through the same frustration: job boards flooded with aggregator listing pages ("500+ DevOps jobs in Pakistan"), senior roles asking for 8+ years when they have 2-3, and remote postings that quietly exclude candidates based in Pakistan.
I couldn't write the backend myself from scratch, but I could design the experience and, with AI as a genuine collaborator throughout the build, put together something that actually solves this. The result is the AI DevOps Job Matcher — an autonomous, serverless career agent that searches the live web for DevOps/Cloud/SRE roles in Pakistan, filters out the listing-page noise, and uses AI to score each real job against a candidate's actual profile — skills matched, skills missing, a realistic match percentage, and a tailored apply strategy. My friends open one page and see only jobs worth their time.
How I Built It: Process, Key Decisions, and Challenges
As a designer working outside my usual domain, I leaned heavily on AI throughout — not just for the "smart" scoring layer, but to help me reason through the AWS architecture itself, since serverless backend design isn't where I normally live.
Challenge 1: Aggregator spam. Early searches kept returning listing pages instead of individual jobs — titles like "144 Azure DevOps jobs in Pakistan." We built a filtering layer that catches these patterns (numeric-prefixed titles, known search-result URL structures) before they ever reach the AI, so we're not wasting analysis on garbage.
Challenge 2: AI defaulting to the same score. Early on, the model gave almost every job an "85% match" regardless of fit, and occasionally invented salary numbers. Getting this right meant being very explicit in the prompt — penalize roles needing 5+ years for a mid-level profile, and never state a salary unless it's actually in the posting.
Challenge 3: Unstable job IDs. We initially used Python's built-in hash() for deduplication in DynamoDB, and jobs kept reappearing as "new" on later runs. Switching to hashlib.sha256 for a stable, deterministic ID fixed it.
AWS Services Used / Architecture Overview
- Amazon EventBridge Scheduler — triggers the agent automatically every few hours, no manual step
- AWS Lambda (Python 3.12) — orchestrates the web search, filtering, and AI scoring
- Generative AI (via OpenRouter) — the "Deep Research Agent" that evaluates each job and returns structured scoring, matched/missing skills, and apply strategy
- Amazon DynamoDB — stores processed jobs with a 2-day TTL so the dashboard only ever shows fresh, currently-active listings
- Amazon API Gateway + Amazon S3 — API Gateway serves the data, S3 hosts the static dashboard my friends actually use
- AWS IAM + CloudWatch — least-privilege access and visibility into each run
What I Learned Across the Summer
Coming from design, the biggest shift was realizing how much of "AI quality" is actually data quality — the filtering layer mattered as much as the AI prompt itself. I also learned firsthand why deterministic IDs matter in distributed systems, and how something as small as a DynamoDB reserved keyword can quietly break an entire response. But honestly, the most rewarding part wasn't the architecture — it was watching this go from "an idea to help my friends" to something they're actually using to find real jobs in Lahore, Islamabad, and Karachi.
Link to App :
Shoutout
Thanks to my friends in Lahore's DevOps community for being the actual reason this exists — this was built for you. And a nod to Lewis Sawe's "Museum That Grows," whose always-on, no-human-in-the-loop design was the inspiration for making this agent run completely on its own.
Loading article