Terminal / source context
Developer tools
CodeGraph CLI
Explore a codebase through its structure, dependencies, and source-grounded answers.
Python · tree-sitter · NetworkX
Muhammad Ali Nasir AI engineer
I build AI systems.
From model to product.
Local AI tools, code intelligence, and applied research — built in Islamabad, Pakistan.
Scroll to exploreSelected work
Engineering, in practice.
Explore the systems, decisions, and evidence behind each project.
6 case studies 1 research study
Terminal / source context
Council / evidence
Knowledge systems
Research Council
A four-agent GraphRAG workspace for reviewing literature and tracing the evidence.
GraphRAG · Four specialist agents
Models / routing
Model infrastructure
LocalForge
A self-hosted control plane for local models, routing, memory, and knowledge workflows.
FastAPI · GGUF · Qdrant
More things I’ve built.
Applied ML, optimization, and agent workflows
Forecasting / coolingClimate ML
S2Cool
Solar forecasting and cooling optimization across four Pakistani cities.
Route / fuel stopsOptimization
Fuel Route Planner
Plan fuel stops around route, tank capacity, station prices, and detour costs.
Intake / proposal workflowAgent workflows
Proposal Bot
Turn discovery transcripts into proposals, with review and revision inside Slack.
Pricing intelligence
Decision Frame
Demand forecasts, causal price elasticity, and bounded pricing recommendations for e-commerce teams.
Applied research
Research, with
practical intent.
Exploring what efficient models can do for languages and problems that deserve more attention.
Explore the research
Language / efficient models
LightUHope
A compact transformer for Urdu hope-speech classification.
- 0.92
- Macro F1
- 3.2M
- Parameters
- 24,124
- Dataset samples
Results reported in the research manuscript. Under review at Springer Nature.
The problem
Identify hope-speech in Urdu while working with a language that has fewer NLP resources and models.
The approach
Build a compact Urdu transformer with 3.2 million parameters, keeping the model practical to train and evaluate.
Evaluation
The manuscript reports 0.92 macro F1 on a 24,124-sample dataset. Model size is 97% smaller by parameter count than mBERT.
Experience
A practice shaped
by building.
2025–2026
Associate AI Engineer DevRolin · Part-time
Transformer fine-tuning, production feature pipelines, RAG systems, and FastAPI services.
2025–2026
Freelance ML Research Engineer S2Cool
Solar irradiance prediction and cooling optimization across 485,000+ climate records.
Research & open source
Independent engineering & research
CodeGraph CLI, LightUHope, and tools for local AI, evidence retrieval, and developer workflows.
About me
The person behind
the systems.
I’m Muhammad Ali Nasir, an AI engineer based in Islamabad, Pakistan. I work across model deployment, graph-based retrieval, and tools that make AI systems easier to use.
My projects range from local-model routing and codebase intelligence to Urdu language research and solar forecasting.
Good ideas deserve
to become useful.
Keep in touch
Let’s talk about your project.
AI systems, developer tools, or a research question. I’d like to hear about it.
muhammadalinasir00786@gmail.com
M. Ali Nasir