Python & AI
Engineer
I build scalable REST APIs, AI-powered RAG systems, and automated AI grading pipelines using Python, FastAPI, LangChain, and LangGraph — with production experience integrating ZKTeco biometric hardware for real clients.
About
who I amI'm Muhammad Subhan, Python & AI Engineer based in Lahore, Pakistan, with over a year of production experience in FastAPI and Django backends, biometric hardware integration, RAG systems, and n8n workflow automation.
At PAC College of Accountancy, I built an AI pipeline that grades scanned exam scripts against structured rubrics and returns an annotated PDF with a verifiable CSV breakdown. Before that, at Mozzine Technologies, I shipped a desktop biometric attendance system for real clients, integrating ZKTeco fingerprint, face, palm, and RFID devices, and supported those clients directly for the full engagement.
I care about reliability, clean architecture, and direct client support. If you need dependable backend systems, RAG chatbots, or automation — let's talk.
Technical stack
tools I work withAI & LLM
Backend
Automation
Biometrics
Databases
Frontend
DevOps
Tools
Work experience
where I've shipped- Built an AI exam-grading pipeline that extracts marking schemes and student answer scripts from scanned PDFs using vision-based LLM extraction, grades them against structured rubrics, and returns an annotated PDF with a verifiable CSV breakdown.
- Engineered a MongoDB-backed data pipeline for model answers, student submissions, and grades, with a reproducible two-pass rubric extraction and an own-figure (OF) dependency graph, so students keep method-mark credit when an earlier figure in their working is wrong.
- Built a marker frontend to upload a student script, select a saved marking scheme, and download the graded PDF with a per-criterion CSV for audit.
- Added deterministic post-processing guards around the LLM grading step (duplicate-credit prevention, sign verification, evidence-column disambiguation) so the saved mark, not the raw model reply, is the source of truth.
- Built a desktop biometric attendance application using Python and PyQt6 (MVC architecture), integrating ZKTeco devices with SQLite3 for attendance data processing.
- Developed secure FastAPI REST APIs for device communication, authentication, and data synchronization.
- Integrated multiple ZKTeco device models (fingerprint, face recognition, palm recognition, RFID card readers) and implemented multi-threaded sync workers (QThread) to keep the UI responsive.
- Provided direct customer support for the full one-year engagement: diagnosed and resolved recognition errors, connectivity, enrollment, sync failures, and deployment issues, and handled day-to-day client queries and escalations.
- Created n8n workflow automations to streamline operational tasks and orchestrate data pipelines.
- Taught 9th and 10th grade for 3.5 years, preparing topic-wise study material and mentoring students on exams, assignments, and problem-solving.
- Tutored online for 1 year via Zoom and Google Meet, with live classes and one-on-one doubt-clearing sessions.
- Years of explaining difficult topics step by step is why I can walk both clients and non-technical stakeholders through technical work clearly.
Selected projects
things I've builtAI RAG Agents & Chatbots
Resolves repetitive support queries with an LLM and vector search over a curated knowledge base. n8n pipelines ingest new documents and FAQs without redeploying the backend, and retrieval is tuned for relevance versus latency.
Users upload PDFs and documents and ask questions in natural language for policy and manual lookups. Chunking and embedding pipelines keep retrieval accurate across large multi-page files.
The same architecture extended for pre-sales and technical queries, giving sales and support teams instant, source-grounded answers.
Aggregates databases, external APIs, and documents with LangChain and LangGraph tool calls into one cited answer. Handles conflicting data between sources and missing records.
General-purpose chatbots with persistent conversation memory on web and WhatsApp, using prompt engineering and context-window management to stay coherent across sessions.
Document AI & Hardware Systems
Grades scanned answer scripts against structured rubrics and returns an annotated PDF plus a per-criterion CSV, with own-figure method-mark handling and deterministic guards on the LLM output.
Desktop application used by real clients, with ZKTeco fingerprint, face, palm, and RFID integration, MVC architecture, FastAPI sync APIs, and QThread sync workers.
Domain Bots & n8n Automation
Topic-wise Q&A over curated healthcare content using retrieval-augmented generation.
Chatbot for CSS and PPSC exam preparation with topic-wise Q&A over curated subject material.
Form submissions flow into the CRM with auto-generated welcome emails, and scheduled or webhook-triggered syncs keep Google Sheets, Airtable, and PostgreSQL consistent.
Monitors API endpoints for uptime and performance and sends Slack and email alerts on failures, with retry and escalation logic to avoid noisy alerts.