Musharaf Hussain Abid, AI Engineer

01 / About

I think like a researcher. I ship like an engineer.

I'm Musharaf Hussain Abid, an AI Engineer with three years of hands-on experience across AI engineering and research roles, building, optimizing and deploying production-grade systems. I specialize in Large Language Models, Retrieval-Augmented Generation, Computer Vision and end-to-end workflow automation. And I bridge the gap between cutting-edge research and scalable enterprise execution.

There's no agency behind this site. When you hire me, you get me: one accountable engineer from the first scoping call to the final handover.

a.

I own the full stack

Architecture, data pipelines, modelling, deployment, monitoring. I take responsibility for the whole system, not just the model. No hand-offs.

b.

I communicate like a colleague, not a vendor

You'll get written updates, weekly demos and honest estimates from me, so you always know exactly where your project stands.

c.

I don't do demos that die in production

A model isn't done when it trains. It's done when it serves traffic reliably and moves a business metric. I hold my work to that bar.

d.

I protect your unit economics

I favor open-source models, caching and right-sized infrastructure, so what I build stays cheap to run long after I hand it over.

02 / Expertise

What I do, end to end.

I've deliberately spent three years covering every layer of the AI lifecycle, so nothing about your system ever becomes someone else's problem.

GenAI & LLMs

RAG architectures·Agentic workflows·Fine-tuning (Qwen, LLaMA)·Prompt engineering·Evaluation & guardrails

Machine & Deep Learning

PyTorch·NumPy·Pandas·Scikit-learn·Experiment design & evaluation

Computer Vision & NLP

Image classification·Object detection·OpenCV·Text analytics·Document understanding

Workflow Automation

n8n·LangGraph·Model Context Protocol (MCP)·API integrations·Human-in-the-loop design

Cloud & AI Infrastructure

AWS·Azure·RunPod·Modal·Replicate·Docker·CI/CD

Databases & Vector Stores

PostgreSQL·MongoDB·ChromaDB·Pinecone·Hybrid & semantic retrieval

03 / Selected Work

Work I put my name on.

Client systems, open-source releases and automation agents, from enterprise LLM pipelines to real-time vision. Each one is judged by the outcome it delivered, not the technology used.

01

Contract-Intelligence Pipeline, in Production

A live enterprise client's construction contracts and project correspondence demanded more than human reading speed. I designed and built a production LLM-based contract-intelligence pipeline that extracts structured data from contract documents and classifies inbound correspondence against standardized clause logic, driving automated compliance alerts. A hybrid semantic-search correlation system (vector search, BM25, neural reranking) links related contractual events across the project's full email and document history, tuned against real production logs rather than fixed heuristics. Every LLM-generated date, monetary value and duration passes a deterministic verification layer, so financial and temporal accuracy never depends on model-generated text.

ClientConfidential · construction enterprise · via Nexora AI (remote)

ScopeArchitecture · build · optimization · production operations

StackPython·Azure OpenAI·Azure Service Bus·Azure AI Search·Azure Blob Storage

Discuss a similar build →

Outcomes

61%

Summary latency cut · 116s to 45s via parallel execution

100%

LLM output behind a deterministic verification layer (dates, values, durations)

Live

In production · real correspondence traffic · zero silent data loss

02

Contract-Event Email Classification

For the same enterprise client, I built a second system: an email classification module that reads inbound project correspondence and routes it automatically across the contract workflow. The classification model handles the full inbound stream and assigns each email to its contract-event category: early warning, meeting minutes, compensation event, quotation, project manager instruction, daily site report, programme health email, or defects notice.

ClientConfidential · construction enterprise · via Nexora AI (remote)

SignalsEmail subject·email body·attachment contents

StackPython·Azure OpenAI·Azure Service Bus·Azure Blob Storage

Automate your correspondence →

Outcomes

8

Contract-event categories classified automatically

3

Signals per email · subject, body & attachments

Live

Feeding automated compliance alerts in production

03

Customer Churn Prediction

Customers rarely announce they are leaving: by the time an account cancels, the decision was made weeks earlier. In this client engagement I fine-tuned machine learning models for customer churn prediction, turning the client's own account data into an early read on the accounts most likely to leave, so the business could act while there was still time to act. Work covered the full modeling path: preparing the data, fine-tuning the models, and evaluating them until the predictions were dependable enough to base decisions on.

ClientConfidential · via Neural Technologies

ScopeData preparation · model fine-tuning · evaluation

StackPython·classical ML

Scope a similar model →

Outcomes

Early read

Accounts most likely to leave, flagged before they do

Fine-tuned

Models shaped on the client's own account history

End to end

From raw account data to dependable predictions

04

Real-Time Crowd Counting System

Dense crowds break off-the-shelf detectors: people stack on people and occlusion eats accuracy. So I fine-tuned YOLOv8n on the CrowdHuman dataset, 15,000 training images of exactly those conditions, and wrapped the model in a real-time application: webcam, image and video inference with live counting, adjustable confidence, and automatic visual and sound alerts when a configurable crowd threshold is crossed. Built for surveillance and public-safety use, released open source.

ContextOpen-source project · public-safety surveillance

ScopeDataset preparation · fine-tuning · evaluation · real-time application

StackPython·Ultralytics YOLOv8n·OpenCV·Streamlit

Discuss a vision system →

Outcomes

0.776

mAP50 · dense & occluded crowds after fine-tune

82.8%

Precision on held-out evaluation

15,000

CrowdHuman images used for fine-tuning

05

WhatsApp CRM Assistant: Conversational Orchestrator

Business requests arrive over WhatsApp as casual messages and get lost by the dozen. I built an LLM orchestrator that reads each incoming message, extracts what the sender actually wants, and routes it to the right sub-workflow: projects, leads or reminders. Conversations hold state across turns through a session state machine, a parallel step-by-step menu mode offers a guided alternative to free text, and a scheduled agent delivers reminders on time. The data layer was migrated to an external ERP API as the single source of truth.

FocusConversational AI · LLM orchestration

ScopeIntent routing · stateful multi-turn flows · scheduled delivery

Stackn8n·OpenAI·WhatsApp Cloud API·PostgreSQL

Automate a workflow like this →

Outcomes

3

Sub-workflows orchestrated: projects, leads, reminders

Stateful

Multi-turn conversations via a session state machine

ERP

External API promoted to the single source of truth

06

Supplier Deals Pipeline: AI Price-Comparison Agent

Sourcing across international marketplaces means manual searching, translation guesswork and prices in mixed currencies. I automated the entire sweep: the pipeline searches six international marketplaces, uses an LLM to translate buyer queries into Chinese and Turkish, normalizes every price to USD automatically, and writes the comparison report itself. A Next.js dashboard presents the results behind strict relevance filtering, so loosely-matching listings never reach the buyer.

FocusAutonomous research agent · e-commerce sourcing

ScopeMarketplace search · LLM translation · report generation

Stackn8n·Apify·OpenAI·Next.js·Vercel

Automate your sourcing →

Outcomes

6

International marketplaces searched automatically

2

Languages handled by LLM translation: Chinese, Turkish

USD

Every price normalized before comparison

07

ImgX: Serverless GPU Image-Enhancement API

Photo-restoration models are heavy, and serving them reliably behind an API is an infrastructure problem as much as a model problem. I deployed GFPGAN and Real-ESRGAN on NVIDIA A10G GPUs behind a serverless Modal endpoint, then stress-tested the service and resolved four production bugs: a pipe-buffer deadlock, a Triton configuration type mismatch, and CUDA out-of-memory failures on high-resolution inputs among them.

FocusGPU inference · image restoration

ScopeDeployment · load testing · production debugging

StackModal·NVIDIA Triton·GFPGAN·Real-ESRGAN

Ship a GPU API →

Outcomes

2

Restoration models served: GFPGAN & Real-ESRGAN

A10G

NVIDIA GPUs behind a serverless API endpoint

4

Production bugs resolved under stress testing

04 / Colleagues & Teams

What the people I work with say.

Worked with me?

If we have worked together and you would like to leave a few words about the experience, you can do it here. I read and verify every submission personally before anything appears on this page.

05 / Services

Engagements I take on.

Whether you need a fixed-scope sprint or a long-term partner, I work the same way: production readiness is the bar on everything I ship.

Custom GenAI & RAG Development

I build production retrieval pipelines over your private data: parsing, chunking strategy, hybrid search, reranking, and grounded answers with citations.

From messy PDFs and wikis to an assistant your team actually trusts.

Typical scope2 – 6 weeks

AI Automation & Agentic Pipelines

I design n8n, LangGraph and MCP workflows that connect your tools, make decisions and eliminate manual work around the clock.

Process mapping, agent design, guardrails and a clean handover to your team.

Typical scope1 – 4 weeks

Fine-Tuning & Deploying Open-Source LLMs

I fine-tune Qwen, LLaMA and Mistral models on your domain data, then quantize, benchmark and serve them cost-efficiently on GPU cloud.

When API costs or data residency rules make open weights the right call.

Typical scope2 – 5 weeks

Computer Vision & NLP Architecture

I architect detection, classification, OCR and text-analytics systems for accuracy, latency and real-time production traffic.

Feasibility through to serving, with evaluation harnesses built in from day one.

Typical scope3 – 8 weeks

Cloud Deployment & Vector DB Integration

I ship Dockerized, CI/CD-driven deployments on AWS, Azure, RunPod and Modal, with ChromaDB or Pinecone powering fast, relevant retrieval.

Reproducible infrastructure your team can operate after I hand over the keys.

Typical scope1 – 3 weeks

How I work: every project, no exceptions

Step 01

Scoping call

We start with a free 30 minutes: your goals, your constraints, and my honest feasibility read, even if my answer is “don’t build this”.

Step 02

Written proposal

I send you a fixed scope, timeline and price with milestones, in your inbox within 48 hours of our call.

Step 03

Build sprints

I demo progress every week and send async written updates against measurable checkpoints. No surprises.

Step 04

Handover

You get full documentation, an evaluation report and a live walkthrough, plus 30 days of post-launch support from me, at no charge.

06 / Questions

Before you ask.

Fixed-scope or hourly?+

I work fixed-scope by default: you approve a defined deliverable, timeline and price before any work starts. Hourly and monthly retainers are available for ongoing advisory and iterative builds.

Who owns the code and the models?+

You do, with full IP transfer on final payment. I regularly work under NDA, and confidential engagements are a normal part of my workflow. Case studies are only published with written permission.

Can you work inside our existing team or codebase?+

Yes. I slot into in-house engineering teams, inherit existing repos, and adapt to your tooling and review standards. For larger builds I bring trusted collaborators, but you always keep a single point of accountability.

What happens after delivery?+

I close every project with documentation, an evaluation report and a live walkthrough. Thirty days of post-launch support is included; ongoing retainers are available if you want me to keep iterating.

How do we start?+

Send me a short note about your project: the data or workflow involved, and your timeline. I reply to every serious inquiry within 24 hours; if it looks like a fit, we book a scoping call and you receive my written proposal within 48.

07 / Contact

Let's build something intelligent together.

Tell me about your data, your workflow, or the system you need built. I'm now booking for Q4 2026 and I take on a limited number of engagements at a time, so every client gets my full attention. I reply to every serious inquiry within 24 hours.

Email Me Book a 30-min intro call ↗ Download CV (PDF) ↓

Islamabad, Pakistan · GMT+5 · working overlap with EU mornings & US East afternoons

musharaf-hussain-abid ↗

+92 323 548 4551 ↗

Direct

Email ↗