SAMIYA SHAHZAD MALIK / APPLIED AI ENGINEER
I connect models, data, and thoughtful engineering to turn practical problems into useful products.
An interactive rules experiment
Direct buyer + clear budget + defined timeline.
Direct buyer: 50 points. Confirmed budget: 30. Defined timeline: 20. A known distributor can earn up to 50; an unknown buyer type requires review.
80–100: high priority. 50–79: follow up. Below 50: gather context. Unknown signals stay unknown.
This is a simplified, browser-only example of rule-based scoring, inspired by LeadScout. It does not call an LLM or reproduce the full project’s scoring rubric.
Illustrative rules. Runs entirely in your browser.
01 / SELECTED WORK
A few things I’ve worked on.
Open a project to see the decisions behind it.
Skills exist before the right words do.
Mapping informal experience to occupational categories, so valuable skills become easier to describe and discover.
AI with a sense of place.
A skincare information prototype connecting image understanding and product discovery with Pakistani users and local availability.
A score you can explain.
A lead qualification API combining language-model extraction, structured validation, and deterministic rules.
A page-prioritization experiment where the baseline outperformed the more complex models. Reporting that result was part of the work.
One experiment. A useful reminder to test the baseline.
02 / THE PERSON BEHIND THE SYSTEMS
The part I enjoy most is taking a vague problem and making something real enough to test.
More of my work on LinkedInI’m an applied AI engineer based in Pakistan. Most of my learning happens through projects: connecting models to workflows, untangling failures, and improving the software around them.
I work across retrieval, agent systems, automation, and machine learning evaluation. I care about useful context, inspectable decisions, and knowing when a person should stay in control.
03 / WHAT’S NEXT?
REMOTE ROLES & PROJECT COLLABORATIONSHave a practical problem and room for thoughtful engineering?
Find me on LinkedIn ↗01 / UNMAPPED
A retrieval and agent workflow for mapping informal experience to ISCO-08 occupational categories.
People can have valuable skills without the formal language to describe them. UNMAPPED explores how AI can help connect that experience to a shared occupational vocabulary.
FastAPI, LangGraph, and ChromaDB connect retrieval with streamed responses. The application was packaged with Docker for Hugging Face Spaces.
Selected for the Top Projects Showcase in the Global HackNation World Bank track. This is a hackathon project, with further retrieval and evaluation improvements to explore.
02 / SKINSIGHT
A skincare information prototype designed around Pakistani users and product availability.
A useful product recommendation needs to account for what someone can actually find locally. SkinSight connects image and ingredient information with that local context.
React and TypeScript provide the interface. Gemini supports image understanding and structured scan results, while Google Search grounding helps connect responses with Pakistani stores and products.
Those structured results also become context for follow-up chat, so the conversation can refer back to the latest scan.
03 / LEADSCOUT
A backend qualification system for turning free-form inquiries into structured signals, priority, and explanations.
The language model extracts buying signals. Structured validation checks that output. Deterministic rules then calculate the score, keeping the decision inspectable.
Evaluation scenarios include prompt injection, missing fields, and distinguishing distributors from end customers.
04 / FLYRANK
Machine learning internship work on prioritizing declining web pages with search and analytics data.
I compared logistic regression and random forests with a baseline for ranking pages. In the final capstone evaluation, the baseline captured 2,131 missed clicks, compared with 1,698 for random forest and 1,274 for logistic regression.
Split design and leakage control matter as much as the choice of algorithm. An evaluation should reflect how the system will actually be used, and a simple baseline deserves a fair comparison.
Internship completed in August 2026.