Muhammad
Shahbaz Sarwar Awan
Full Stack & AI Engineer, Web Applications, LLM Systems & RAG Pipelines
AI graduate from FAST (NUCES) who builds complete products, not just models. I work across the stack: React and Next.js on the front end, FastAPI and Node.js services on PostgreSQL behind it, and the LLM layer that ties it all together. Comfortable with RAG pipelines, agentic workflows, embeddings and vector search, grounded generation, and containerized deployment. I've shipped AI automation for a US healthcare client and delivered production chatbots and workflow automation for international clients.
Products shipped end to end
A mix of production AI systems, client work, and independent research, spanning voice agents, RAG pipelines, full-stack platforms, and computer vision.
Aeromek Call Agent
Browser-native voice AI for live lead qualification
A single-page voice interface that lets a visitor talk to a conversational AI agent right in the browser, with nothing to install and nothing to download. The agent runs a structured discovery conversation, figures out which AI solution the visitor needs, and captures contact details for follow-up in real time.
- —Real-time browser audio capture streamed to a speech-enabled LLM over a live channel
- —Permission-gated call widget with a live transcript panel
- —Structured discovery flow that classifies intent (voice, sales, support, booking, chat, custom)
Aeromek Video Studio
Prompt-to-video generation studio with AI narrators
An AI-powered studio that turns a single natural-language prompt into a finished, voice-over-driven vertical video. The pipeline plans a narrative, casts a narrator persona, storyboards scenes, sources or generates b-roll, and renders a captioned 1080×1920 MP4 with lip-synced voice-over and music.
- —Automated pipeline: concept planning → scene blueprint → b-roll → render
- —Character and voice-casting subsystem with lip-synced narration
- —Storyboard-based scene editor for manual fine-tuning
WhatsApp RAG Assistant
Autonomous WhatsApp agent with retrieval-grounded replies
A production-oriented WhatsApp automation that drives WhatsApp Web to detect unread one-to-one chats, filters out groups and channels, and replies using a multi-provider LLM chain augmented by a locally hosted RAG pipeline that auto-reindexes as new documents are added.
- —Selenium-driven session control with layered heuristics to isolate real 1:1 conversations
- —Multi-provider LLM fallback chain (Gemini, Groq, OpenAI) for reliability
- —Auto-reindexing Chroma vector store built from a watched documents folder
RAG Bot
Conversational Q&A over any PDF, grounded in its content
A document question-answering system that lets a user upload a PDF and ask natural-language questions about it. Documents are chunked and embedded, and every answer is generated from the most relevant retrieved passages rather than model memory alone.
- —FastAPI backend for ingestion, chunking, and vector retrieval
- —Streamlit conversational frontend with per-session document scoping
- —HuggingFace sentence-transformer embeddings with ChromaDB storage
Mortgage Brokerage CRM
Multi-tenant SaaS CRM for a lead-to-disbursement sales pipeline
A full-stack CRM built for a mortgage brokerage to run its Telesales and Sales teams end to end. Leads captured from Meta, website, WhatsApp, TikTok, and referral channels are qualified by Telesales, then automatically handed over as a Case to Sales, which pushes each one through bank submission, approval, and disbursement.
- —Role-based multi-tenant architecture with server-enforced permissions per role
- —Lead pipeline with duplicate detection, callback scheduling, and a full activity timeline
- —Case pipeline tracking multi-bank submissions through to disbursement, with a live sales dashboard
Plant Disease Detection
YOLOv8 computer vision for real-time crop disease identification
A computer vision system that detects and classifies plant leaf diseases from images and video in real time, trained on ~2,600 annotated leaf images spanning apple, corn, potato, and tomato crops, and deployed through a browser-based interface.
- —Anchor-free single-stage YOLOv8 detection architecture
- —Frame-level object tracking for video input
- —Dataset curated and annotated from Roboflow Universe sources
Counting-Sheep
Vision-based livestock counting under uncontrolled camera conditions
A computer vision system that automates detection, tracking, counting, and segmentation of sheep in video footage, built to work with handheld and third-party footage rather than a single fixed industrial camera. That makes it a much harder problem than standard counting setups.
- —Two-stage architecture: adjustable region-of-interest delineation + YOLO detection/tracking
- —Instance segmentation for accurate counts in dense, overlapping groups
- —Designed to generalize across varied, uncontrolled camera angles
Potato Counter
Real-time conveyor-belt produce counting with YOLOv11
An automated counting system for food-processing lines that detects, tracks, and counts potatoes moving along a conveyor belt using a tracking-by-detection paradigm anchored to a user-defined polygonal counting zone.
- —YOLOv11 detection combined with a polygon-zone counting strategy
- —Tracking-by-detection to avoid double-counting on a moving belt
- —Built for fixed-camera industrial deployment
HealPay: AI Medical Billing Automation
Final Year Project, graded A+
A full-stack platform that ingests medical documents, extracts clinical data with OCR and NLP, recommends billing codes with LLMs, and validates claims before submission. It's the same category of pipeline I later shipped in production for a US healthcare client.
- —OCR + NLP extraction of demographics, diagnoses, and procedures from unstructured records
- —LLM-based ICD-10, CPT, and HCPCS code recommendation with supporting evidence
- —Claim validation against payer requirements ahead of submission
Financial Recommender System
Fine-tuned LLM recommendations with traceable sources
A recommendation system built on an LLM fine-tuned on financial data and paired with a retrieval layer, so every recommendation the model produces can be traced back to the source data it came from.
- —LLM fine-tuning on domain-specific financial data
- —Retrieval layer for source-attributed recommendations
- —Built for explainability, not just accuracy
Architecture lab
Illustrative breakdowns of three real pipelines above: how a query, a call, or a lead actually moves through the system, step by step.
The shape behind RAG Bot and the WhatsApp RAG Assistant: a document set gets chunked and embedded once, then every incoming question is answered from the passages that actually match it, not from model memory alone.
Where I've built
Production engineering roles delivering RAG platforms, voice agents, and full-stack systems.
AI Engineer, Paisol Technology
- —Owned the full build of Chetti, an internal platform that turns a single prompt into a deployable RAG chatbot, embeddable with one line of iframe code. That cut setup from days of developer integration down to minutes of self-serve configuration.
- —Engineered the full retrieval layer, including ingestion, chunking strategy, embedding storage, and grounded generation with source attribution, and measured it against a hand-built domain test set.
- —Chose open-source LLMs over hosted APIs to keep per-deployment inference cost flat and client documents in-house, tuning chunk size and retrieval depth to work within that constraint.
- —Built a multi-turn booking assistant with explicit conversation state management, integrated with Google services for scheduling.
AI Engineer, Aeromek AI
- —Replaced manual chart review with an automated clinical document pipeline that pulls demographics, diagnoses, and procedures out of unstructured medical records using OCR and NLP.
- —Implemented LLM-based ICD-10, CPT, and HCPCS code recommendation returning ranked, evidence-backed candidates, moving human coders from manual lookup to quick validation.
- —Built FastAPI microservices on PostgreSQL that check claims against payer requirements before submission, flagging non-compliant records ahead of rejection.
Software & AI Engineer, Independent Contract Work
- —Delivered retrieval-augmented assistants over client document sets as the sole engineer, handling ingestion, chunking, retrieval tuning, front-end integration, deployment, and iteration.
- —Standardized a reusable embeddable chatbot pattern across engagements, replacing bespoke per-client builds with configuration-driven deployment.
- —Connected LLM reasoning to clients' CRM, inbox, and scheduling systems through n8n workflows and REST integrations.
Engineering toolbox
Production-proven tools and frameworks used across web apps and AI pipelines.
Frontend
Backend
AI & LLMs
Frameworks & Tools
Data & Retrieval
Cloud & DevOps
FAST National University of Computer & Emerging Sciences
Graduated 2026BS in Artificial Intelligence
Coursework in Agentic AI, NLP, MLOps, Machine Learning, Recommender Systems, Databases, and Web Technologies.
Let's build something worth shipping
Available for all roles and contract work: full-stack builds, AI/LLM systems, or a bit of both. Reach out and I'll get back to you.