pro
Fariha Muazzam
I build production-ready AI systems that actually ship
New to Contra
Fariha is ready for their next project!
Satellite imagery is the ultimate stress test for Computer Vision.
It is not just about detection; it is about architecting precision across massive, high-resolution datasets where every pixel counts.
I have been building AI pipeline to segment parking lots, walkways, and structures from aerial data. The challenge is not just the mode, it is the Infrastructure:
- Tiling & Inference: Processing massive dimensions without losing small-object context.
- Robust Pre-processing: Using OpenCV to handle varying lighting and atmospheric "noise."
- Cloud Orchestration: Turning heavy segmentation models into responsive, scale-ready tools.
It is about turning raw pixels into a reliable data layer for the business.
Currently opening 20 hours/week for long-term partnerships in Computer Vision & AI Systems.
1
54
AI-based Video Generation Pipeline
0
1
I designed and delivered an end-to-end MVP for a real-time AI talking avatar system, where live browser speech is converted into an AI-generated response and rendered through a lip-synced avatar video on GPU hardware.
The objective was to validate technical feasibility, latency characteristics, and perceived real-time interaction before committing to production hardening. The system integrates speech-to-text, LLM-based reasoning, text-to-speech, and video synthesis into a single, runnable pipeline, deployed on an A100 GPU.
0
256
At Formulatrix, I led the development of a computer vision pipeline for RockMaker, a biotech product used in crystallization experiments. I built and deployed object detection and image classification models that improved scoring accuracy by 13%, reducing manual effort for scientists and increasing customer satisfaction. The solution was productionized with Docker and CI/CD pipelines, ensuring scalability and reliability across client sites.
0
217
This project showcases a working prototype of a Legal AI Risk Analyzer built using Retrieval-Augmented Generation (RAG), FAISS vector search, and a local Llama3 model.
The system ingests contract documents, breaks them into contextual clauses, retrieves semantically similar clauses from a vector database, and uses the LLM to assess risk levels, reasoning, and recommendations.
Designed entirely for local deployment, it ensures data privacy while demonstrating how enterprise teams can analyze legal and compliance risks without exposing sensitive content to external APIs.
0
166
Problem:
Manual video production pipelines don’t scale and break when content length or structure changes.
What I Built:
An agentic workflow where LLMs:
Decompose a brief into structured scenes
Generate per-scene scripts, visuals, and metadata
Maintain consistency across scenes
Output machine-readable payloads for downstream video assembly
Why it matters:
Handles variable-length scripts (10–50+ scenes)
Designed for automation, not prompt hacking
Built for integration into larger systems
0
191
Problem:
Manual tracking of opportunities and prioritization wastes time and lacks consistency.
What I Built:
An automated workflow that:
Monitors inbound signals (emails etc)
Extracts and normalizes structured data
Uses LLM logic to prioritize and tag opportunities
Logs everything into a searchable system
This pattern generalizes to:
lead qualification
ops automation
AI-assisted decision pipelines
0
193