Building full-stack
products end-to-end.

Built the AI hiring stack for Preplify, the company's HR management platform. An LLM résumé-parsing stage turns raw CV text into structured candidate profiles; screening scores each applicant against the job description per criterion — skills match, experience relevance, education fit, career stability — instead of one opaque number; shortlisted candidates sit a structured AI-conducted first interview that's transcribed and scored the same way. HR can override every automated decision. Scenario analysis put HR's hands-on time for a 120-applicant opening at ~1.75 hrs, down from ~34.5.

I also benchmarked the model options behind it — self-hosted Qwen2.5-1.5B via Ollama and a fine-tuned Qwen3-0.6B against hosted OpenAI APIs — weighing accuracy against cost-per-résumé and keeping candidate PII in-house, then shipped browser prototypes of the hiring panel and parser flow and presented the recommendations to the team.

A free, offline-first personal finance and receipt manager built for Pakistani users — three codebases written solo: a Flutter mobile app, a Node/Postgres REST API, and a React admin panel. On-device OCR receipt scanning, expense and income tracking, budgets, subscriptions, grocery price comparison, and rewards, running on free-tier infrastructure at zero monthly cost.

The version that matters is the one that stopped asking people to type anything. Qeemat reads bank and wallet alerts through an Android notification listener and parses them into transactions entirely on-device — message text never leaves the phone. The patterns are deliberately loose, because Pakistani banks change their wording without notice, and anything the parser isn't confident about is queued for a one-tap confirmation instead of quietly landing in the ledger.

Around that sits the part a finance app can't skip: a biometric + PIN app lock with attempt lockout, screenshot blocking, an encrypted local cache keyed to the Android Keystore, 15-minute access tokens with rotating refresh tokens and replay detection, and account deletion that really cascades. Multi-account balances, transfers that are never counted as spending, udhaar tracking, household members, and monthly reports sit on versioned SQL migrations, covered by 80+ tests running in CI against a real Postgres. Offline writes queue in an outbox and replay in order, idempotently — verified in airplane mode.

Inside the app

— scroll →

Splash → onboarding → sign-in → dashboard

Onboarding

Insights

Capture consent

Bank alert

It landed