AI Agentic Developer & LLM EngineerAvailable for new work

Muhammad Faraz Khan

I build multi-agent systems that move millions of records and act on them. So far that's 10 systems shipped, the largest handling 1.7M records, and three products of my own that are live today: ScholarReach, Mail Scheduler and VibeUI.

About

About me

I'm a computer science graduate from NED University who fell for what happens when large language models meet real production data. At Marktec I pulled 1.7 million leads out of MySQL, Excel and Gmail into one PostgreSQL database, kept it in sync in real time with Kafka and Debezium, and put a multi-agent system on top that writes personalised outreach at scale.

A Vision-LLM resume analyzer, a RAG assistant over 500+ documents, an ML system forecasting climate extremes: different problems, same habit. I build AI that ships.

Education
BS Computer ScienceNED University (NED-UIT)Sep 2021 – Jul 2025
Based in
Karachi, PakistanOpen to remote and on-site rolesAvailable now
Certifications
HCCDA-AI · HuaweiData Science Bootcamp · DataCrumbsDatabases · Coursera / Meta
Focus
Agentic AI · RAG · LLM appsData engineering · CDC pipelinesML and computer vision

Experience

Where I'veworked

Feb 2026 — PresentNow

Marktec

AI Agentic Developer

Lead Management & Email Automation Platform — Client: DigitEMB (1.7M leads)

  • Ingested 1.7M leads from MySQL, Excel and Gmail into a unified PostgreSQL database.
  • Built real-time Change Data Capture with Apache Kafka + Debezium to continuously sync all sources.
  • Designed a deduplication & normalization pipeline resolving every lead to one canonical record.
  • Integrated Reoon Email Verifier for deliverability scoring and an OpenAI-powered enrichment bot for missing attributes.
  • Architected a multi-agent system (OpenAI Agents SDK): a Retrieval Agent queries PostgreSQL, a Copywriter Agent generates personalized outreach.

Python / PostgreSQL / Apache Kafka / Debezium / OpenAI Agents SDK / Docker

The pipeline I built at Marktec

1.7M records · 2 agents · live

Lead pipeline, in order, as data flows left to right: sources feed PostgreSQL, then dedupe and normalize, then verify and enrich, then a Retrieval Agent and Copywriter Agent generate and send outreach.

ingest

MySQL · Excel · Gmail

PostgreSQL

Dedupe + Normalize

validate

Reoon Verify

AI Enrich

agents

Retrieval Agent

Copywriter Agent

delivered

Outreach Sent

Jan 2026 — Feb 2026

DataCrumbs

AI Engineer Intern

Resume Analyzer & Email Outreach Agent

  • Built a Streamlit app accepting resumes (image/PDF) + job descriptions, auto-saving uploads to Google Drive.
  • Implemented a Vision LLM pipeline extracting skills, experience & layout — scoring resume-to-JD match (0–100) with structured feedback.
  • Developed an email agent that formats & delivers the full analysis report to candidates; deployed on Streamlit Cloud.
  • Shipped an n8n workflow sending personalized daily proposal emails from Google Sheets — zero manual intervention.

Python / Streamlit / Vision LLM / Google Drive API / n8n / Google Sheets API

Projects

ThingsI've built

Four I designed, built and shipped on my own, start to finish. Smaller builds follow below.
ScholarReach landing page: 'Reach the professors who work on exactly what you want to research.'

FlagshipSolo full-stack engineer

ScholarReach

Full-stack web app · For students applying abroad

Finds professors who actually research your niche, gets their email addresses, and sends personal, reviewed outreach emails from your own Gmail.

Students applying for BS research, MS and PhD positions abroad were contacting professors at random — browsing university sites one by one, copying names into spreadsheets, guessing emails and sending generic messages to people who didn't work on their topic, so most got no reply. ScholarReach turns weeks of manual searching into minutes and keeps every email personal and approved by the student.

Demo available; private installs on request (VPS or own device).

Highlights

  • Seven specialized agents run it end to end — profile, matching, email finding, department scouting, positions, writing and sending
  • Matching Agent ranks professors from recent OpenAlex papers by topic match, lab leadership, funding and activity — every match shows its reasons
  • Email Finder validates each address against the professor's name and university domain; likely-format guesses are clearly marked unverified
  • Writer Agent drafts a personal email citing the professor's own paper; Sender Agent sends from the student's own Gmail (OAuth, send-only) only after approval
  • Long scrapes split into ~20s resumable steps stored in the DB, so they survive serverless timeouts
  • Responsible scraping: respects robots.txt, blocks requests to private addresses (SSRF), honors 429/503 Retry-After
  • Security: bcrypt, JWT invalidated on password change, account lockout, single-use reset links, encrypted Gmail tokens

Stack

Backend
Python / FastAPI / SQLAlchemy (async) / PostgreSQL (Neon) / BeautifulSoup / httpx
Frontend
Next.js 16 / React 19 / TypeScript / Tailwind CSS 4
Integrations
Gmail API / OpenAlex / Europe PMC / ORCID / Google Gemini / Tavily / Serper
Hosting & quality
Vercel / Neon (Frankfurt) / Cloudflare DNS / 90 pytest tests

Office hours, by recipient

You, Karachi

London

New York

Sydney

your clock
0006121824

Each email goes out inside its recipient's office hours, not yours.

NewestSolo: product, backend, MCP integration, frontend and deployment

Mail Scheduler

AI agents · MCP · Automation · Full-stack · 2026

Client emails that arrive in their office hours, even with your laptop closed.

Freelancers and agencies emailing clients across time zones send at the wrong time: their working hours are the client's night. Gmail's Schedule send works one email at a time in your own time zone, and AI assistants' Gmail connectors can only send immediately, not later.

Mail Scheduler is an MCP-powered fix. Connect Gmail once and add a personal connector URL to Claude, ChatGPT, Cursor or any MCP client. Attach a client sheet in chat and describe the email; the assistant shows a dry-run preview of every rendered email with its send time in the client's local time, and schedules only after you confirm. A server queue then sends each one from your own Gmail inside the recipient's office hours, while your devices are off.

Built and shipped end to end in a few days, with Claude Code as pair programmer. Runs entirely on free tiers.

Highlights

  • Per-recipient office hours: each client's IANA time zone, working hours and workdays, with weekends and late hours rolling to the next opening
  • Works from any MCP client — Claude, ChatGPT (developer mode), Claude Code, Cursor — or from a web dashboard without AI
  • Mail-merge personalisation from sheet columns ({first_name}, {company}…); a row is rejected if a placeholder is missing
  • Engagement signals ranked by reliability — reply detection over IMAP (headers only), bounce detection, click and open tracking — with security-scanner and preview-bot hits filtered out, because a tracking pixel alone is unreliable under Apple Mail Privacy Protection and image blocking
  • Multiple sending addresses per account and multi-user accounts; atomic claim-before-send so nothing goes out twice, retries with backoff, and safe handling of removed senders
  • Remote MCP server on serverless: a stateless MCP app is built per request, with no long-lived session manager or startup events
  • AI Gmail connectors send immediately and can't reach Gmail's scheduled send, so sending lives in a server-side queue, which is what makes it work on every AI platform. An hourly Claude cloud routine was the first prototype; cron-triggered SMTP replaced it for minute-level timing

Stack

Backend
Python / FastAPI / SQLAlchemy / PostgreSQL (Neon)
AI integration
Model Context Protocol / Python MCP SDK v2 / Streamable HTTP (stateless)
Email
SMTP / IMAP
Dashboard & hosting
Vanilla JS / Vercel (serverless) / cron-job.org / Cloudflare DNS

A call, end to end

  1. Caller

    phone line

  2. Vapi

    speech in and out

  3. Agents SDK

    Gemini

  4. FastAPI

    validation

  5. PostgreSQL

    patient record

Voice AISole developer

Voice AI Patient Registration Agent

Voice AI agent · Patient intake · 2026

A phone agent that registers new patients through natural conversation, validates every detail, and saves the record to a database with a REST API and dashboard.

Callers dial a real U.S. phone number and talk to an AI intake coordinator. The agent collects name, date of birth, sex, phone, address and optional insurance and emergency-contact details, in whatever order the caller gives them. It re-asks for invalid values such as a future birth date or a six-digit phone number, accepts spelled-out corrections, reads everything back, and only saves after the caller confirms.

Behind it is a FastAPI service: Vapi handles the phone line and speech, the OpenAI Agents SDK runs the conversation on Gemini, and the agent's tools call the same validated service layer as a REST API. Records live in PostgreSQL and are visible on a dashboard alongside call transcripts and booked appointments.

Deployed on Railway during development; the hosted demo runs on trial credits and may be offline.

Highlights

  • Real U.S. phone number answered by the agent — Vapi handles telephony, speech-to-text and text-to-speech
  • Out-of-order conversation: callers can give several details at once, correct anything mid-call, or switch to Spanish
  • Field-specific re-prompts for invalid data — future birth dates, wrong-length phone numbers, invalid states or ZIP codes
  • Confirmation enforced in code, not the prompt: the save tool first returns needs_confirmation, so a read-back and caller reply are guaranteed before any write
  • One validation layer shared by the agent's tools and the REST endpoints, with the key rules repeated as database CHECK constraints
  • The LLM runs in the backend as an OpenAI-compatible custom-LLM endpoint, so prompt, tools and model choice stay version-controlled
  • Model fallback when one is rate-limited, idempotent tool actions so a retried turn can't duplicate a patient, and a spoken apology if every model fails
  • Returning-caller detection by phone number, appointment booking, stored transcripts, and 22 automated tests across the API, agent actions, LLM endpoint and call webhook

Stack

Backend
Python / FastAPI / SQLAlchemy / Pydantic
Agent & voice
OpenAI Agents SDK / Google Gemini / Vapi / Deepgram
Data
PostgreSQL / SQLite
Dashboard & delivery
HTML/JavaScript / Docker / Railway / 22 pytest tests
VibeUI landing page: 'Turn a conversation into a real UI component'

Dev toolSolo full-stack / AI engineer

VibeUI

Full-stack AI web app · Developer tool

AI-powered UI generation from a conversation — describe your product, get a full set of production-ready React screens grounded in real competitor research.

VibeUI turns a short 4-question chat into a complete, multi-screen product UI. It interviews the user, researches their real competitors live on the web, and generates three production-ready React + Tailwind screens (e.g. landing + dashboard + settings) rendered in a live in-browser sandbox — built for people who can write the logic but freeze on frontend design decisions.

Highlights

  • Conversational 4-question “vibe interview” that produces a structured design brief
  • Live web-grounded competitor research (Google Search via Gemini) with a real-time activity log
  • Multi-screen generation — 3 model-chosen screens per product, not a single component
  • Live sandboxed preview (Sandpack) with code view and a screen switcher
  • Multi-tenant auth with encrypted, per-user “bring your own key” API keys
  • Per-account project history, plus a selectable generation model (Gemini 2.5 Flash / Pro)

Stack

Frontend
Next.js 16 / React 19 / TypeScript / Tailwind CSS v4 / Zustand / Framer Motion / Sandpack
Backend
FastAPI / Google Gemini 2.5 / Google Search Grounding / Fernet Encryption
Data & Auth
Supabase / PostgreSQL / Row-Level Security
Deployment
Vercel / Custom Domains

Skills

What Iwork with

AI & LLMs

  • OpenAI Agents SDK
  • LangChain
  • RAG
  • Prompt Engineering
  • GPT-4
  • Gemini
  • LLaMA 2
  • Vision LLM

Data Engineering

  • PostgreSQL
  • MySQL
  • Apache Kafka
  • Debezium
  • ChromaDB
  • FAISS
  • Pinecone

Programming & Frameworks

  • Python
  • Flask
  • Streamlit
  • TensorFlow
  • Keras

Tools & Platforms

  • Docker
  • Git
  • n8n
  • REST APIs
  • Google Sheets API
  • AWS

Testimonials

In theirwords

Worked with Faraz on a data scraping project, and he did an excellent job. He gave me exactly what I wanted and was very flexible whenever I needed changes. My client on this project was a bit demanding, giving instructions day by day, but Faraz handled it well and met every deadline my client set. The leads he provided were good quality too. It was a really smooth experience overall, and I'm looking forward to working with him long term!

Kay · Founder & CEO, Konnex

“Faraz took ownership of a 1.7-million-lead migration for our client and built the entire CDC pipeline and multi-agent outreach system with minimal supervision. He works like an engineer twice his experience level.”

Abdul Mueed · Team Lead, Marktec

“During his internship, Faraz shipped a complete Vision-LLM resume analyzer end-to-end — from pipeline to deployment — in weeks. He learns fast and delivers production-quality work.”

Abis Hussain · AI Mentor, DataCrumbs

“Faraz's final-year project on climate-extreme prediction showed strong applied ML skills and genuine research discipline. One of the more self-driven students I've supervised.”

Dr. Lubaid · FYP Supervisor, NED University (NED-UIT)

Contact

Got a project?Let's talk.

A project in mind, a role to fill, or just want to talk agents and RAG? Send me a message. I usually reply within a day.