Open to new-grad & junior roles · Islamabad, PK · CS graduate, 2026

I build AI automationthat runs in production— not demos.

I turn high-volume operational work — inbox triage, review response, candidate evaluation, creative generation — into autonomous, event-driven systems. Full-stack, end to end: OAuth and API integrations, background-job pipelines, cost-aware LLM generation, and the operator dashboards that keep humans in control.

View the work Download résumé

Selected evidence

6–10×

faster large-review syncs

Documented · Revvy PERFORMANCE.md

50–200ms

cached page loads

Documented · Revvy benchmark

30–50%

lower generation cost, pre-filtered

Target · design target

engineers led at Genesys Research Lab

Team lead · Jun–Aug 2025

Core stack

Next.jsTypeScriptPython · FastAPIPostgreSQL · PrismaOpenAI · Anthropic · GooglePinecone · WeaviateAWSDockerpg-boss · Inngest

01 — Selected work

Systems built to run unattended

A focused set of projects, each with a clear problem, an architecture decision worth defending, and an honest account of what shipped. Private and client work is described by domain, scale, and role — never by customer data.

01RevvyPrivate productA multi-tenant SaaS that ingests Google Business Profile reviews at scale and generates on-brand reply drafts through a tunable draft-to-approve-to-publish pipeline, engineered so throughput and per-call cost stay controlled once real review volume arrives.Next.js 14 (App Router)TypeScriptReact 18PostgreSQLPrisma ORM

02Emmy — thread-aware Gmail categorizationClient workA production Gmail auto-categorization system that routes known senders by deterministic rule and classifies the rest with a thread-aware LLM, writing every decision back as a native Gmail label — and logging why.Next.js 14 (App Router)TypeScriptPrisma / PostgreSQLInngest (durable background jobs)Gmail API + Pub/Sub push

03Recruitment RAG PlatformInternal labAn end-to-end recruitment platform that matches candidates to roles by semantic retrieval and runs a conversational, non-repeating interview agent — built with a 4-person team I led during a research-lab internship.PythonFastAPIWeaviateNomic embeddingsLlama 3 8B (Groq API)

04AdzeePrivate productA solo, end-to-end AI product that turns a brand profile and a one-line theme into a full ad set — brief, ranked headlines, copy, and imagery — through a staged, brand-conditioned pipeline with a self-correcting image-quality loop.Next.js 15React 19TypeScriptOpenAI GPT-4oOpenAI image generation

See all work

02 — How I work

Production thinking, not just prompts

The model is usually the last decision, not the first. Most of the engineering is everything around it.

Integrate at the source

Real OAuth token lifecycles with refresh buffers and invalid-grant detection; webhooks and Pub/Sub for near-real-time ingestion — not happy-path API calls.

Engineer for throughput and cost

Rate-limit-safe concurrency, batched database writes, tagged caching, and pre-filtering that skips work before it costs a model call.

Keep humans in control

Draft-to-approve pipelines, structured reasoning logs, confidence heuristics, and per-decision auditability — automation you can actually trust to run.

03 — Beyond solo builds

Team leadership and research depth

I led a four-person AI research team at Genesys Research Lab, delivering a RAG-based recruitment evaluation platform on lab infrastructure.

Separately, I contributed across a production multi-tenant agentic AI SaaS platform — multi-tenant billing, multi-model routing, an AWS pipeline, and vector search — as one engineer on a larger team.

I co-built a multimodal music-recommendation system as my Computer Science thesis, and I'm the technical founder of an early-stage startup building evidence-grounded CAD intelligence — with the discipline to know when not to build.

More about how I work