Rayyan Mohsin — AI Engineer

About

I'm happiest turning rough ideas into tools people can actually use.

My work ranges from agents that manage ad campaigns to vision systems that analyse fight footage and pipelines that turn prompts into finished videos. I care less about a flashy demo than whether the thing still works when someone depends on it.

Pakistan · UAE

Proof of Work

View all →

01 / 06Active

40+ Tools

LLM agent

Meta Ads Manager

Facebook/Instagram ad management run through an LLM agent — it creates campaigns, generates creatives, and adjusts performance over the Meta Graph API. The hard problems turned out to be the boring ones: OAuth, token encryption, rate limits, and getting 40+ tools to behave.

02 / 06Live

<60s

render time

Launchable

Describe your product in chat and it comes back as a 1080p demo video about a minute later — an agent writes the scenes, picks the type and motion, and keeps everything on brand. It started as a question: how much of a video editor can one good agent replace?

03 / 06Live

9 Concepts

from one URL

Halftone

Paste a product URL and it reads the page into a brand kit — audience, offer, palette, tone — then stages nine ad concepts on a canvas, rendering images on demand. Most of the work is in the guardrails: type, color, and crop rules enforced at generation time, and budget caps so a session can't quietly burn money.

04 / 06Client Work

3-Model Pipeline

computer vision

TenCount

Fight footage in, punch analytics out. Three models chained together: fine-tuned YOLOv11m tracks the fighters, YOLOv8m reads their pose, and an AttentionBiLSTM classifies six punch types from the motion. Built to replace hours of manual fight review.

05 / 06Live

3 Sources

Reddit · YouTube · web

Agent Memory

Research agents that scrape Reddit, YouTube, and the web, then file what they find. Everything stays searchable in plain English through RAG over vector embeddings — so nothing your team already learned gets researched twice.

06 / 06Live

Side Project

for peers

CV Maker

A no-friction CV builder I shipped on a weekend for friends and juniors at university applying for their first internships and jobs. Pick a template, fill the form, export a clean PDF — no signup, no paywall. Built because the existing tools felt hostile to first-time job seekers.

Internal Tooling

Client intelligence · ClientAcquisition.io

Internal

The Brain

Our client roster lived only on the platform where we talked to clients — nowhere queryable. I turned it into a data pipeline feeding an optimized Obsidian clone: one force-directed graph — every client a node, sized by contract, colored by health. Click a node and the whole file opens. A Hetzner-hosted agentic layer re-parses and normalizes the data daily, so it never goes stale.

On top sits a RAG agent: the team asks in plain English — “which clients are at risk and why?” — and gets a grounded answer instead of digging through files. It reshaped how the team works: account reviews take minutes instead of combing scattered notes, everyone operates from the same up-to-date picture, and at-risk clients surface before they become fires — a more organized, more efficient workflow.

191

Clients, one graph

Daily

Agentic refresh

RAG

Ask, don't dig

UGC Ad Pipeline

Nothing in this video was filmed. A pipeline turns one photo and a short voice note into a finished, post-ready vertical ad — cloning the voice, driving a talking-head, generating the b-roll, and assembling the captions, cutaways, and timing automatically. Each cut then gets an automated review pass before export.

This clip is the real output, built from a single headshot and a 40-second voice sample.

Autonomous Media

Two brands · one pipeline

Live

Content Engine

@footiepulse and @9thneuron are two social accounts I built to run with minimal hands-on work. Each day, the system finds timely stories, turns them into short videos with voiceover and captions, and publishes them to Instagram and TikTok.

One engine powers both accounts while keeping their subjects, voice, and visual identity distinct. The result is a repeatable content operation that can move from a trending topic to a finished post without rebuilding the process for every brand.

2

Brands, one codebase

37

Posts shipped

$0

Voice + score, local