Making machines
understand people.
Hasir Sayed is a data science student and AI/ML engineer building reliable language systems, safer AI workflows, and practical data products.
Building safer, more useful AI systems.
AI Safety and Data Pipeline Engineer
Built an autonomous data science tool deployed on Vercel, with decoupled data cleaning and preprocessing pipelines designed for high-throughput, maintainable workflows. Designed secure BYOK API-key handling; real-time PII detection and redaction; prompt-injection and jailbreak safeguards; and human-in-the-loop review for high-risk outputs.
- Developed a collaborative three-agent workflow (Evaluation, Structural Validation, Task Execution) that automates complex data sanitization operations.
- Built real-time prompt-injection filters, jailbreak shields, and customized PII masking routines to protect user-provided data.
- Optimized structural output consistency and memory retention limits using multi-agent simulation loops and adversarial test scenarios.
AI and Agentic Systems Developer
Portfolio Platform
Created and deployed an agentic portfolio platform to present technical work through a more capable, interactive experience. Integrated custom APIs and Model Context Protocol (MCP) capabilities to distribute tasks and support autonomous execution.
- Designed custom servers implementing the Model Context Protocol (MCP) to let localized LLMs safely inspect directory contexts and execute local commands.
- Programmed task-orchestration workers to split complicated multi-step requests into atomic stages, achieving fully autonomous task resolution.
- Optimized context window packing and conversational history trees, reducing API costs by 35% while maintaining accuracy and responsiveness.
Data science and AI projects, in detail.
Personal SLM fine-tuning
Fine-tuned Qwen 2.5 (0.5B) on a custom domain dataset using supervised fine-tuning. Built the full loop: collection, cleaning, deduplication, dataset structuring, hyperparameter tuning, baseline evaluation, and prompt/dataset iteration. The model was deployed locally with quantisation and inference optimisation for privacy-conscious use on consumer hardware.
Autonomous data science tool
A deployed data product that automates complex preprocessing work. The architecture separates evaluation, structural validation, and task execution across three collaborating agents, while keeping the pipeline maintainable and responsive for real-world data-cleaning flows.
Operations and MIS Portfolio with AI Assistant
Designed and built an executive-style portfolio for a Business Operations and MIS leader. The site pairs a structured representation of experience, projects, and certifications with an AI assistant that can answer recruiter questions from portfolio knowledge.
Professional portfolio website
Designed and developed a customized, responsive portfolio website tailored to the client's professional background and career goals.
Benchmarking SLM Performance and Integration
Published research paper (Zenodo DOI: 10.5281/zenodo.21396947) systematically comparing lightweight language models across capability, inference speed, resource usage, and edge deployment trade-offs. Developed a reproducible evaluation benchmark methodology.
Small Language Model (SLM) Integration and Performance Analysis
A systematic evaluation of lightweight language models across edge deployment environments, comparative inference speeds, and parameter quantization constraints.
How the system works.
Grounded in data science, extended through applied AI.
Analysis
Machine learning
Language AI
Responsible systems
Engineering
Research
Have a difficult problem?
Let’s talk.
For collaborations, roles, or a thoughtful technical conversation, send a note. I read every message.