shifa mehreen
back

Projects

Systems, models, and experiments I've shipped along the way.

VISTA

Spatio-temporal RAG for video

A persistent memory layer for frontier models that understands egocentric video — tracking objects, spatial orientation, and temporal context through allocentric heading and scene graphs. Outperformed Gemini 3 Flash and Molmo-2 on spatial-temporal reasoning.

Gemini Flash · OpenCV · RAG · Agentic AI

Cloud Orchestrator

Natural language → live infrastructure

A multi-agent system that translates plain English into executable GCP infrastructure DAGs across 20+ services (Vertex AI, GKE, BigQuery), cutting cloud setup time 95% through automated tool orchestration.

Gemini 2.5 Flash · ADK · GCP · Multi-Agent

FeedForward

Multilingual AI for food banks

A voice, chat, and web platform that automates form processing for food-bank operations — multilingual, sub-2s latency, built on OpenAI APIs with Twilio integration.

GPT-4o · LangChain · Flask · Twilio · AWS RDS

Olympiad Reasoner

Math reasoning, fine-tuned

Fine-tuned 5 LLMs with LoRA on 10K ReAct+ToRA samples for National Olympiad-level math, with a weighted self-consistency ensemble that climbed the leaderboard 350 positions.

Qwen 2.5 · DeepSeek R1 · LoRA · ReAct · ToRA