LLM & agentic systems
Assistants, RAG and multi-model workflows that cite their sources, return structured output and hold up under real traffic.
- OpenAI
- Claude
- LangChain
- RAG
- Pinecone
- FAISS
- Agents
I build AI that sees, reasons and repairs.
I’m 
Muhammad Imran — AI engineer, published researcher and Co-Founder & CEO of PipelineMender AI. I’ve worked with Trilogy on Ephor.ai and, over 5+ years, shipped LLM, computer-vision and data systems for teams across Europe, North America and Asia-Pacific.
pipelinemender / incident PM-1042
MonitoringCopy_Orders · Microsoft Fabricrun 8f2a·c41
Failure
Investigation
Evidence
Root cause
Fix
Illustrative run · see PipelineMender AI
Experience across
TrilogyPipelineMender AIephor.aiCM SERVIZICloudquikNESL LLCfreelancerMDPI ElectronicsBahria UniversityIIU Islamabad
- Years shipping AI & software to productionSince 2021
- 5+
- Average rating across 38 client reviewsFreelancer.com
- 4.9/5
- On-time and on-budget delivery recordFreelancer.com
- 100%
- Citations of my first-author surveyGoogle Scholar
- ~100
Open to design partnersMuhammad Imran
AI Engineer & Founder
- Based in
- Islamabad, Pakistan
- Local time
- UTC+5
01About
Research rigor. Founder speed.
I’m an AI engineer who moves models out of notebooks and into products people rely on — LLM assistants, document intelligence, computer-vision pipelines and the full-stack platforms around them.
My path runs from a BS in Software Engineering and an MS in Data Science — with a thesis on underwater image classification — to a first-author MDPI survey cited ~100 times, then engineering roles with Trilogy, Cloudquik and NESL, and contract work for CM SERVIZI in Rome.
Today I’m Co-Founder & CEO of PipelineMender AI, where failed data-pipeline runs become evidence-backed root causes and reviewable fixes. Alongside it, I still partner with teams on hard AI problems.
- Focus
- LLM systems · Computer vision · Data reliability
- Working with
- Teams in Europe, North America & APAC
- Education
- MS Data Science · BS Software Engineering
01
Production over prototypes
Evaluation, monitoring and failure modes are part of the first commit — not an afterthought.
02
Evidence before answers
Systems that show their work, with every conclusion traceable to its source.
03
Private by design
From on-prem LLMs to read-only connectors, sensitive data stays where it belongs.
04
Clear communication
Plain-language updates, honest trade-offs and delivery on the date we agreed.
02Now building
Turning failed pipelines into reviewable fixes.
When a data pipeline breaks at 2 AM, the error tells you where it stopped — not why. PipelineMender AI investigates failed runs on Microsoft Fabric, Azure Data Factory and Databricks, connects the evidence and hands your team a specific fix to review.
PipelineMenderAI
Co-Founder & CEO · since Oct 2026Your pipeline failed. Here’s the fix.
- Watches every 5 minutes
- Read-only connections sync on a schedule; a failed run becomes an incident automatically.
- Evidence, not guesses
- Logs, activity output, definitions and schema become numbered evidence behind each root cause.
- A fix you can review
- A specific SQL, code, configuration or credential change — never applied without you.
- Private by default
- Scoped, read-only access. Customer pipeline data is never used to train models.
Co-founded with Anza Abbasi (COO).
How an investigation runs
- 01
Failure
A failed run is detected and opened as an incident with its run context.
- 02
Investigation
Logs, activity output, pipeline definition and schema are read.
- 03
Evidence
Each finding is kept as numbered evidence you can inspect.
- 04
Root cause
The likely cause, its confidence and the evidence behind it.
- 05
Fix
A specific change, ready for your review in your own tenant.
Built forMicrosoft FabricAzure Data FactoryDatabricks
Schema driftExpired credentialsMissing input filesSource timeoutsDependent task failuresType mismatchesConfiguration errorsPermission changes
03Expertise
End-to-end AI engineering, notebook to production.
I work across the whole stack, so the model, the data pipeline and the product around it are designed together — not handed between teams.
Private & on-prem AI
Local LLMs, document intelligence and SSO on your own hardware — the data never leaves the building.
- Ollama
- Open WebUI
- Docling
- Qwen
- IIS
- OIDC
Computer vision
Classification, detection, measurement and OCR — from an MS thesis on underwater image classification to deployed APIs.
- YOLOv4
- OpenCV
- CNNs
- Transfer learning
- CLAHE
- Tesseract
Data platforms & MLOps
Pipelines, lakehouses and model serving on Microsoft Fabric, Databricks and AWS.
- Fabric
- Data Factory
- Databricks
- Lambda
- Step Functions
- ECS
Full-stack product engineering
The product around the model: APIs, dashboards, PWAs, auth and the infrastructure to run them.
- React
- Next.js
- TypeScript
- FastAPI
- PostgreSQL
- Redis
Security & IoT
Hardened access, encrypted tunnels and device telemetry for privacy-sensitive environments.
- OIDC SSO
- WireGuard
- ThingsBoard
- OWASP
- GrapheneOS
Daily toolbox
PythonTypeScriptReactNext.jsFastAPIFlaskPostgreSQLRedisDockerOpenAIAnthropicGeminiLangChainHugging FaceOllamaPyTorchTensorFlowOpenCVscikit-learnNumPypandasJupyterDatabricksAWSAzureNode.jsTailwind CSSLinuxWireGuardVercel
04Selected work
Selected work, shipped to production.
A cross-section of recent engagements — multi-LLM platforms, on-prem AI, LLM products, research-grade models and the operational software around them.
LLM platform
Ephor.ai — multi-LLM research platform
Worked with Trilogy · Austin, TX
Worked with Trilogy on Ephor.ai, building AI-first full-stack features for a platform where multiple LLMs collaborate on research, knowledge management and workflow automation.
- Agentic workflows and multi-model LLM integrations
- React + TypeScript frontend on FastAPI, PostgreSQL and Redis
- Shipped end to end in a globally distributed team
- React
- TypeScript
- FastAPI
- PostgreSQL
- Redis
- Docker
On-prem AI
Private AI platform on Windows Server
CM SERVIZI S.r.l. · Rome
Self-hosted document intelligence, private chat and single sign-on — every model runs on the client’s own hardware.
- PDF table extraction & analysis API with Docling and Ollama
- Open WebUI over HTTPS behind IIS, with OIDC SSO extended to legacy Classic ASP / VB6 apps
- ThingsBoard IoT with an FTP telemetry bridge and WireGuard tunnels
- FastAPI
- Ollama
- Docling
- Open WebUI
- IIS
- OIDC
- ThingsBoard
LLM product
AI-first universal inbox
Product build
One inbox for Gmail, Microsoft 365 and IMAP — with AI thread summaries, streaming reply drafts and inbox prioritisation.
- Gmail OAuth, Microsoft Graph and IMAP (Yahoo, AOL) behind one interface
- Server-side Claude and OpenAI with Zod-validated structured outputs
- Installable PWA with service worker and offline mode
- TypeScript
- Claude
- OpenAI
- Zod
- Microsoft Graph
- PWA
Local LLMs
Private AI coding server
Developer client · Italy
A turnkey VirtualBox appliance that gives a developer a private, ChatGPT-style coding assistant — no command line needed.
- Ollama + Open WebUI as auto-starting Windows services that survive reboots
- Offline Qwen2.5-Coder 7B plus cloud models answering side by side
- RAG knowledge base over the developer’s own documents and source code
- Ollama
- Open WebUI
- Qwen2.5-Coder
- RAG
- Python
- VirtualBox
Speech + NLP
Multimodal emotion recognition
Research build
A conversational emotion model on the MELD dataset that fuses what is said with how it is said — and remembers earlier turns.
- RoBERTa-large (text) + Wav2Vec2 (audio) with cross-modal attention
- Speaker-aware gated fusion and a contextual transformer over dialogue history
- Focal loss for class imbalance and layer-wise LR decay, trained on an A100
- PyTorch
- RoBERTa
- Wav2Vec2
- Transformers
- Colab A100
Operations software
Car-park operations platform
BOI Carpark · New Zealand
Day-to-day operations for a vehicle-storage yard — invoices, a visual key box, long-term contracts and scheduled billing in one place.
- Live yard occupancy computed from key-slot usage
- Visual key box with long-term locker slots and key assignment
- Scheduled account billing, cash-up and end-of-day reporting
- JavaScript
- Web app
- Scheduled billing
- Reporting
05Computer vision
Computer vision, from thesis to production.
My MS thesis — a CNN that classifies underwater marine species — started years of hands-on work in image classification and object detection. Every project below is public on GitHub.
MS thesis · Bahria University · 2022
“Underwater Image Classification for Marine Using CNN”
A multi-label classifier that recognises three marine species in underwater imagery. I trained a custom CNN, compared five pretrained architectures through transfer learning, tested softmax against binary outputs and measured how image enhancement affects accuracy.
ClassesRockFishTileFishStarFish
- accuracy on two classes
- 99%+
- accuracy on the third
- 97%
- pretrained CNNs compared
- 5
- Custom model
- Convolutional neural network with ReLU activations
- Output layer
- Softmax and binary (sigmoid) variants compared
- Pre-processing
- CLAHE image enhancement vs raw images
- Data
- Underwater marine image dataset
Transfer learning · pretrained architectures compared
- VGG16
- VGG19
- ResNet50
- Xception-71
- DenseNet121Best
06On the web
Product sites and storefronts, designed and deployed.
Beyond AI systems, I design and ship fast, polished web experiences for clients — 30+ builds deployed on Vercel in 2026 alone.
07Experience
Five years. Four continents. Always shipping.
From on-site engineering in Islamabad to remote roles for teams in Austin and Rome — and now building my own company.
Oct 2026 — Present
Islamabad · Remote
CurrentCo-Founder & CEO
PipelineMender AI · Full-time
Building an AI data engineer for failed pipelines. When a run fails on Microsoft Fabric, Azure Data Factory or Databricks, PipelineMender collects the evidence, explains why it broke and drafts the fix — turning hours of log-digging into minutes.
- Leading product, customers and go-to-market; currently onboarding design partners
- Evidence-first investigations: observed facts kept separate from hypotheses, every root cause traceable
- Read-only by design — nothing is changed in customer pipelines without review
- Data Engineering
- Microsoft Fabric
- Azure Data Factory
- Databricks
- LLMs
- Product
Feb 2021 — Present
Remote
Current Independent AI engineering for founders, CTOs and enterprise teams — a verified Preferred Freelancer with a 4.9 rating across 38 reviews and a 100% on-time, on-budget record.
- Delivered AI, ML, computer-vision, LLM and enterprise-automation solutions across industries
- Designed customer-service automation, document intelligence, predictive analytics and process-optimisation systems
- Led discovery, solution architecture, pricing proposals and long-term client engagements
1 more contributionShow less
- Built SaaS products with Python, FastAPI, OpenCV, OpenAI, LangChain, PostgreSQL, AWS and Docker
- Python
- FastAPI
- OpenCV
- LangChain
- OpenAI
- AWS
- Docker
May 2026 — Sep 2026
Rome, Italy · Remote
AI Engineer
CM SERVIZI S.r.l. · Contract
Self-hosted AI, security and IoT on Windows Server — fully on-premise for data privacy.
- PDF Analysis API: FastAPI service extracting and analysing PDF tables with local models (Docling, Ollama)
- Private AI chat: self-hosted Open WebUI with local LLMs, published over HTTPS through IIS
- Secure access: IIS reverse proxy and OIDC single sign-on for the AI chat and legacy Classic ASP / VB6 apps
2 more contributionsShow less
- IoT platform: ThingsBoard + PostgreSQL for device monitoring and firmware updates, with WireGuard tunnels
- Mobile security: device hardening with GrapheneOS
- Python
- FastAPI
- Ollama
- Open WebUI
- IIS
- OIDC
- ThingsBoard
- WireGuard
Jun 2025 — Sep 2025
Austin, Texas · Remote
Worked with Trilogy building AI-first full-stack features for Ephor.ai — a platform for AI-assisted research, knowledge management and multi-LLM collaboration.
- Engineered frontend and backend systems with React, TypeScript, Python, FastAPI, PostgreSQL, Redis and Docker
- Implemented LLM integrations, agentic workflows, API services and data-driven features
- Owned work from architecture to deployment and optimisation in a globally distributed team
- React
- TypeScript
- FastAPI
- PostgreSQL
- Redis
- Generative AI
Apr 2024 — Jul 2025
Islamabad · On-site
Full Stack Engineer
Cloudquik · Full-time
Designed LLM assistants and serverless AI workflows for client-facing and internal automation.
- Built LLM assistants and custom GPTs with OpenAI, LangChain and Pinecone
- Integrated AI workflows with event-driven AWS infrastructure — Lambda, S3 and Step Functions
- Created multimodal tools for transcript analysis, document Q&A and speech scoring
1 more contributionShow less
- Fine-tuned open-source LLMs and deployed them on AWS ECS and Hugging Face Inference Endpoints
- OpenAI
- LangChain
- Pinecone
- AWS Lambda
- Step Functions
- Hugging Face
May 2021 — May 2022
Islamabad · On-site
Full Stack Developer
NESL LLC · Full-time
Shipped AI-first features with GPT APIs, fine-tuned models and vector search.
- Deployed solutions using OpenAI GPT APIs, fine-tuned LLMs and vector search with Pinecone and FAISS
- Built end-to-end pipelines for support automation, internal tooling and intelligent data extraction
- Integrated AI into web platforms with FastAPI, React and serverless APIs
1 more contributionShow less
- Owned architectural decisions, sprint planning and rapid prototyping
- OpenAI
- FAISS
- Pinecone
- FastAPI
- React
- Serverless
08Research & education
Grounded in research.
My MS in Data Science produced a CNN thesis on underwater species classification and a first-author survey in MDPI Electronics, now cited around 100 times on Google Scholar. I bring the same rigor to production work: every claim backed by evidence.
Electronics (MDPI)Peer-reviewed · 2021First author
Comprehensive Survey of IoT, Machine Learning, and Blockchain for Health Care Applications: A Topical Assessment for Pandemic Preparedness, Challenges, and Solutions
Muhammad Imran, Umar Zaman, Imran, Junaid Imtiaz, Muhammad Fayaz, Jeonghwan GwakVol. 10, Issue 20, Article 2501 · 14 October 2021
A topical survey of how IoT, machine learning and blockchain combine in healthcare — from medical-report management and drug traceability to tracking infectious diseases — with an assessment of pandemic preparedness, open challenges and future research directions.
- articles reviewed
- 263
- publisher databases
- 6
- citations (Google Scholar)
- ~100
DOI 10.3390/electronics10202501
Research interests
- Machine learning
- Image classification
- Object detection
- Multimodal learning
- LLMs & agents
- Healthcare AI
- Blockchain
Credentials
GRE General Test
ETS · 2024
IELTS Academic
British Council · 2022
Verified · Preferred Freelancer
Freelancer.com · Current
Verified work history (Trilogy)
LinkedIn · Verified
2020 — 2022
MS, Data Science
Bahria University, Islamabad
Thesis“Underwater Image Classification for Marine Using CNN”
Multi-label classifier for rockfish, tilefish and starfish: a custom CNN plus transfer learning with VGG16, VGG19, ResNet50, Xception-71 and DenseNet121 (best). 99%+ accuracy on two classes and 97% on the third.
Thesis details2014 — 2018
BS, Software Engineering
International Islamic University (IIU), Islamabad
Project“University Student Management System”
Built the university’s Student Registration and Fee Management systems with my team — several modules were deployed by the university.
09Testimonials
Trusted by clients worldwide.
Freelancer.com
- Verified
- Preferred Freelancer
4.9/ 5
Average across 38 verified client reviews
- On time
- 100%
- On budget
- 100%
“Excellent service, nothing was too hard and he did everything to a high standard. Will be hiring for another project.”
Shane · New Zealand
BOI car-park platform
“Muhammad works very fast, he is exceptionally good at communication and his computer vision skills are impressive also.”
Freelancer client · Australia
Pixel-perfect front-end build
“A very knowledgeable and patient consultant. Able to resolve the issues highlighted on the topic of AI.”
Valentino · Italy
Private AI coding server
“Great job! On time, understood the task and delivered it excellently.”
Teo · Germany
Survey web app
“Excellent consultant, expert in many IT areas, and always available.”
Valentino · Italy
GrapheneOS device security
10Contact
Have a hard AI problem? Let’s build the fix.
I’m onboarding design partners for PipelineMender AI and partnering with teams on LLM, computer-vision and data-platform work. Tell me what you’re building.
PipelineMender AI & design partners:imran@pipelinemender.com
Find me online
Islamabad, Pakistan · local time (UTC+5) · working across time zones