Muhammad Umer

AI Engineer specializing in multi-step LLM systems and document intelligence pipelines.

Based in: Lahore, PakistanCurrently: AI & Automation Engineer, SellerplexExperience: 2 years

Muhammad Umer is an AI Engineer with a proven track record of shipping production-grade agentic workflows and financial-filings parsers with high accuracy. He specializes in automating complex supply chain and finance processes using LangChain, n8n, and custom LLM extraction pipelines.

Experience

AI & Automation Engineer

SellerplexRemoteworkMar 2026 – Present

Summary

Built a multi-user agentic freight RFQ system in n8n that runs the full quote lifecycle: Slack-triggered intake, AI-filled shipment details from chat replies and packing lists, dispatch by email or forwarder portals (Playwright), automatic follow-ups, and LLM extraction of quotes from replies and attachments. Cut quoting time by an estimated 25–40%. Built an agentic client-communication workflow that requests transfer orders in client Slack channels and parses replies, even out-of-thread or days later, matching each one to the right shipment and updating the planning sheet. Built and shipped five multi-user automations used across the supply chain team; on one client account, they cut the work from two people to one. Built an Amazon FBA shipment plan automation (SP-API, Slack, Google Sheets/Drive) across three inbound flows (SPD partnered, SPD non-partnered, LTL non-partnered), from plan creation through label delivery. Adopted by the full team for a task that was ~25% of its workload, cutting time spent by 50–55% (per user feedback). Built client-facing Slack alerts that turn SKU-level shipment errors into plain language, and secured Slack tokens behind an authenticated n8n proxy. Automated advance shipping notice (ASN) emails for 11+ suppliers, generated from new shipment plans with per-account templates and attached labels. The team now only reviews and approves, roughly halving ASN time.

Results

On one client account, automations cut the work from two people to one.
Automation adopted by the full team for a task that was ~25% of its workload, cutting time spent by 50–55% (per user feedback).
Roughly halved ASN time for 11+ suppliers.

n8nPlaywrightSlack APIAmazon SP-APIGoogle SheetsGoogle DriveLLMPython

AI Engineer

2IQ ResearchworkMar 2025 – Mar 2026

Summary

Automated an ownership-filings workflow previously handled manually by a 5–6 person team: scrapers fetch financial forms, a custom code-based parser handles the core form, LLMs parse the remaining forms into enforced output schemas, and the extracted records are stored in PostgreSQL. Built that core form parser and validated it during development on 7,000+ documents, with the team manually reviewing its output: 96.5% accuracy. Built an end-to-end pipeline for HKEX share buyback filings: scraper → PDF parser → prompt-driven LLM step (LangChain) that processes the parsed output into structured JSON and CSV for validation, then stores the results in a database. Replaced a fully manual process; still running in production over a year later. Set up Arize Phoenix to trace production LLM calls, collecting traces and cleaning datasets for evaluation. Built an internal multi-agent analytics prototype translating natural language to SQL, with guardrails for cost, latency and failures. Built 12+ production scrapers in n8n across global stock exchanges and financial sites, each running end to end on trigger: extraction, cleaning, schema-enforced validation and database ingestion. Designed PostgreSQL schemas and optimized queries for high-volume financial data storage and analytics.

What I did

Developed a code-based parser for ownership filings validated manually across 7,000+ documents.

Results

Validated core form parser at 96.5% accuracy on 7,000+ documents.
Replaced a fully manual process for HKEX filings; still running in production over a year later.
Automated workflow previously handled by a 5-6 person team.

LLMPostgreSQLLangChainArize Phoenixn8nSQLJSONCSVPDF ParserDeepSeekGemini

Skills

Technical

n8n

LLM Extraction

LangChain

Prompt Engineering

Agentic AI

Python

RAG

Slack API (Bolt)

Amazon SP-API

LangGraph

Playwright

FastAPI

PostgreSQL

Docker

Arize Phoenix

Ollama

Data Augmentation

Input Optimization

Digital Ocean

Gemini

DeepSeek

Projects

Skin-lesion segmentation

Lead Developerpersonal

Trained LB-UNet, a lightweight U-Net model for medical image segmentation, locally and deployed it as a FastAPI service (Uvicorn) in Docker behind Nginx on DigitalOcean. Also run open-source LLMs locally with Ollama.

LB-UNetFastAPIUvicornDockerNginxDigitalOceanOllama

Urdu discourse parsing research

Researcheracademic

Ran input ablation studies (removing punctuation, removing connective words, shuffling the text) and measured the change in performance to investigate how UniRST, a multilingual RST discourse parser never trained on Urdu, is able to split Urdu text into elementary discourse units (EDUs) and build valid discourse trees.

UniRSTPythonNLP

AI lecture assistant

personal

Developed an assistant that researches web topics to generate presentation slides with cited sources and a downloadable PPT file, including a draft review stage for human-in-the-loop interaction.

LangGraph

Human-in-the-loop slide generator

personal

A slide generator that handles web research and source attribution using multi-turn state management.

LangGraph

Education

MS Computer Science

Information Technology University, LahoreComputer Science

Thesis. Graph relations and knowledge representations in RAG — evaluating and improving reliability in multi-hop reasoning

Coursework

Medical Image Computing

BS Information Technology

Virtual University of PakistanInformation Technology

Hire MuhammadReach out about a role, a contract or a conversation.For recruiters

Muhammad's twin is AI, it can make mistakes.

More people on Kanvis

Browse more people Create your own Kanvis