AI & Software Engineer shipping production-grade AI systems — autonomous multi-agent platforms, RAG pipelines, LLM orchestration and AI automations — end to end, from model to cloud infrastructure. Currently building AI-first products at Arbisoft and pursuing an MS in Artificial Intelligence at LUMS.
Years building AI systems
AI projects shipped
Production platforms
MS AI CGPA @ LUMS
// What I offer
From a proof-of-concept model to a fully deployed, audited, autonomous AI platform — I cover the whole stack.
Autonomous agents that reason, use tools and coordinate — classifier-driven orchestrators, subgraph lanes, human-in-the-loop approval gates and full audit trails.
End-to-end business automation: document & email ingestion, structured data extraction, scheduled agent runs, and Slack-based reporting — running unattended overnight.
Retrieval-Augmented Generation over your documents and knowledge bases — chunking, embeddings, vector search, prompt engineering and conversation checkpointing.
Classical ML pipelines from data preprocessing to evaluation — classification, regression, ranking and matching systems tuned with rigorous metrics.
Neural networks built and trained in PyTorch — CNNs, Transformers and custom architectures, with disciplined training loops, normalisation and experiment tracking.
Real-time semantic segmentation, optical flow, perception–control pipelines, text classification, semantic search and document understanding.
Production chat experiences: streaming LLM responses, semantic search over site content, document Q&A, and knowledge-base ingestion with lifecycle management.
Async REST APIs with FastAPI and Django — Pydantic validation, async SQLAlchemy, Alembic migrations, background workers and real-time sessions.
AI systems deployed properly: containerised services on AWS, infrastructure as code, secrets management, private networking and event-driven scaling.
Nightshift Agents — an autonomous multi-agent platform I built that runs unattended overnight, with sensitive actions gated behind human approval and every step audited.
n8n pipelines pull from Google Drive and Gmail into a central data store through a secure gateway.
Claude-powered agents extract structured data and propose work via a custom MCP tool gateway.
Propose-never-self-apply safety model; one-retry-then-escalate control flow; WORM audit logging.
Results and proposed actions land in Slack for a human operator every morning.
// Where I've worked
Production AI systems shipped for real users — not demos.
Built an autonomous multi-agent system running unattended overnight: n8n orchestration, custom MCP gateway on AWS ECS (private VPC, bearer-token auth, field-level encryption, WORM audit logs), agent reasoning on Amazon Bedrock (Claude), and Terraform-provisioned infrastructure.
Architected a production multi-agent system with a classifier-driven orchestrator fanning work across RAG, integration and escalation lanes; Chroma-backed RAG with PostgreSQL-checkpointed conversations; Next.js + FastAPI interface with streaming chat.
Async FastAPI APIs with an OpenAI-powered interview evaluator off-request via Taskiq, LiveKit real-time interview sessions, S3 presigned uploads, and an async role–candidate matching & ranking pipeline.
Call-review web app analysing customer call summaries, with long-running AI inference offloaded to AWS Lambda for event-driven, cost-efficient scaling; deployed on EC2/Lambda/S3.
Led Python 3.12 enablement and deprecation fixes across the Open edX stack, keeping CI green on a large-scale open-source platform and reducing inter-component coupling.
// Things I've built
A selection of AI, ML and deep learning projects across agents, retrieval, vision and NLP.
Production-ready RAG plugin for Wagtail/Django — semantic search and LLM-powered chat over site content, with persistent indexing, metadata filtering and scalable vector retrieval.
End-to-end perception–control from monocular video: PIDNet real-time 6-class segmentation fused with dense optical flow and a PD yaw controller producing heading & speed commands.
End-to-end document question answering: PDF chunking, embeddings, ChromaDB vector store and context retrieval for LLM answers — tuned chunking & retrieval for relevance.
CNN implemented and trained from scratch in PyTorch with preprocessing and normalisation to stabilise training; tracked loss/accuracy across epochs.
TF-IDF + Logistic Regression classifier with full text preprocessing, evaluated with precision, recall and confusion matrix analysis.
Canny edge detector from scratch · image segmentation · movie genre classification · RISC-V processor pipeline — see GitHub & Kaggle for the full list.
// My toolbox
The stack I use to take AI systems from idea to production.
// Background
Lahore University of Management Sciences (LUMS)
CGPA 3.88University of Engineering & Technology, Lahore
CGPA 3.827// Get in touch
Whether it's an autonomous agent platform, a RAG chatbot over your data, an ML model, or automating a workflow end-to-end — I'm open to consulting and freelance engagements.
📧 m.tayyabtq@gmail.com