Humaila Farooq

Built a RAG pipeline that ingests PDFs, chunks and embeds them with OpenAI's text-embedding-3-small, stores vectors in ChromaDB, and answers questions with GPT-4o-mini. Wrote the FastAPI backend the React app talks to — async request handling, session management, and caching so we weren't hitting the LLM for the same query twice. Tuned embedding thresholds and added query routing, which made retrieval noticeably more relevant before results even hit the model.

AI Developer Intern at AIK Solution (2026-02 – 2026-05)

• Built a RAG pipeline that ingests PDFs, chunks and embeds them with OpenAI's text-embedding-3-small, stores vectors in ChromaDB, and answers questions with GPT-4o-mini.
• Wrote the FastAPI backend the React app talks to — async request handling, session management, and caching so we weren't hitting the LLM for the same query twice.
• Tuned embedding thresholds and added query routing, which made retrieval noticeably more relevant before results even hit the model.

AI Pull Request Review Agent at AIK Solution

• Set up a 5-agent review pipeline (Triage, Security, Performance, Maintainability, Judge) that kicks off automatically on every GitHub PR via webhooks.
• Added a triage step so only the agents that actually apply to a diff run — dropped token usage by 60-84% compared to running everything every time.
• Compressed diffs and pulled cross-file context to save another ~70% on tokens; the Judge agent rolls all the individual reviews into one comment on the PR.
• Tested it against 6 cases — config breaks, hardcoded secrets, SQL injection, N+1 queries, duplicate code, and clean PRs — and it got all of them right.

Murabaha AI Assistant — Islamic Finance Chatbot at AIK Solution

• Built a RAG chatbot grounded in AAOIFI Standard No. 08, with the standard split into 67 chunks in ChromaDB via LangChain.
• Gave it two modes (Compliance and Learning), session memory with query rewriting, and a check that flags citations that don't actually exist in the source.
• Scored 96.7% (52/55) on a test set that required exact clause citations, not just a plausible-sounding answer.

Data Science Intern at 10 Pearls Pakistan (2025-09 – 2025-11)

• Built an AQI forecasting pipeline in TensorFlow/Keras, from raw sensor data and missing-value handling through LSTM training and evaluation.
• Used SHAP to show which features actually drove predictions, so results made sense to non-technical stakeholders.
• Helped hook the model into a CI/CD pipeline so it retrains and redeploys automatically as new data comes in.

Bachelor of Computer Science in BSCS – Mohammad Ali Jinnah University (MAJU) (2022-09 – 2026-06)