Humaina ltd. | Case Studies

Anonymer Kunde

Edtech

Automatisierte KI-Inhaltsmoderation für soziale Medien bereitstellen

Herausforderung:

Provide an automated AI content moderation for the public social media profiles of various football clubs.

Herangehensweise:

We fine-tuned BERT for various downstream tasks to multiple categories of hate and abuse. We also fine-tuned RESNET50 for computer vision to detect nudity and "NOT_FIT_FOR-WORK" content in images. Trained YOLO5 objection detection for various hateful objects such as weapons, racist symbols, certain vulgar hand gestures and used Roboflow to automate the pipeline and enable non-technical staff to label and train object detection models.

Lösung:

We built a patent pending content moderation algorithm that identities various categories of abuse in speech and also image video and audio using fine-tuned Transformers. GCP Cloud Infrastructure and Python REST API's using FastAPI and Mongo DB. We then built a containerized FastAPI back-end to serve the models, hiredFront end and infrastructure Engineers during the scaling phase. Implemented testing, logging and cloud security as well as ISO127001 and GDPR compliance.

IBM

Finanzsektor

Migration eines On-Premise-Systems für eine renommierte Retailbank

Herausforderung:

Migrate an on-prem system for a major retail bank to RedHat openshift Kubernetes using Terraform.

Herangehensweise:

We migrated an on-prem system of a well-known retail bank to RedHat openshift Kubernetes using terraform. We took this job because we wanted to learn more about networking and security, which we did, while also learning that we like to build things. We were able to build a Golang tool that interacts with Terratest to test the terraform infra code deployed on the IBM cloud, and used gobuster to pentest it during the infra testing phase.

Lösung:

We built a Golang tool that interacts with Terratest to test Terraform infra code to the IBM cloud and used Gobuster to pentest it during the the infra-testing phase.

ProperHome

Immobilien

Aufbau eines Frameworks für das Scraping von Immobilienpreisen

Herausforderung:

Develop a business intelligence platform that learns real estate prices in Portugal in real time. Develop predictive algorithms to accurately price newly listed properties.

Herangehensweise:

We used a fully automated data ingestion pipeline and machine learning classifiers (Support Vector Machine, Linear Regression, Naive Bayes).

Lösung:

We built a framework for scraping real estate prices that evolved into a full-fledged data pipeline, including training Scikitlearn prediction models to deliver prices to insurance customers on a Flask backend, and a Plotly Dash dashboard for analytics (which we managed to build on top of Flask with Auth using the free version of Plotly, which was a challenge, but we are very proud of the results):

Deutsche Automobil Treuhand

Autoindustrie

Extraktion von Gebrauchtwagenpreisen und -merkmalen in aufstrebenden Märkten

Herausforderung:

Provide automated AI content moderation for the public social media profiles of various football clubs.

Herangehensweise:

We fine-tuned BERT for various downstream tasks to multiple categories of hate and abuse, and fine-tuned RESNET50 for computer vision to detect nudity and "NOT_FIT_FOR-WORK" content in images. We then trained YOLO5 objection detection for various hateful objects such as weapons, racist symbols, certain vulgar hand gestures, and used Roboflow to automate the pipeline and enable non-technical staff to label and train object detection models.

Lösung:

We have built a patent-pending content moderation algorithm that identifies different categories of abuse in speech (NLP) and also in image and video (computer vision) and audio (NLP, speech to text) using fine-tuned transformers. GCP cloud infrastructure and Python REST API's using FastAPI and Mongo DB. We then built a containerised FastAPI backend to serve the models, hired front-end and infrastructure engineers during the scaling phase. Implemented testing, logging and cloud security as well as ISO127001 and GDPR compliance.

Anonymer Kunde

Öffentlicher Sektor

Revolutionierung des Rechtssystems mit einem in Großbritannien ansässigen Kunden

Herausforderung:

Create a CHATGPT competitor that gives real UK legal advice to citizens.

Herangehensweise:

FInetuned a Large language model on British law and served as a ChatGPT like chatbot to give free legal advice to the public. Used falcon 9b open source base model and finetuned by provided legal data.

Lösung:

Used React.js (chat-ui and Svelte kit) to deploy a UI that looks like ChatGPT. Once finetuned the model is portable enough to deploy on kubernetes.

Anonymer Kunde

Raumfahrtsektor

Zusammenführung von zwei Datensätzen mit Sattelitenbildern für verschiedene Raumfahrtagenturen

Herausforderung:

Merge two datasets with different resolutions (9km grid vs 25km grid) and different Coordinate Reference Systems to generate a daily gap free soil moisture product.

Herangehensweise:

Merged ESA CCI and ERA5Land datasets after averaging the former to daily from hourly and applying geospatial data tranformations in multiple dimensions using Xarray and python. Built native Numpy functions to derive various statistical measures that feed into a mathematical formula that interpolates data from cloudy days and provides the most accurate estimate for ESA CCI data that suffers from cloud obstruction Nan values.

Lösung:

Built a data pipeline for merging two datasets of sattelite images for different space agencies using Python, including projections and devised an algorithm generating time series of 22 years of data to obtain real time observations to fill in cloudy days. Deployed automated Data Pipeline including the new merging algorithm using Dagster in Python.

Sprich mit unseren KI-Experten über dein Unternehmen

Wir verfügen über umfassendes Know-how bei der Strategieentwicklung, der Entwicklung von Proof of Concepts und der anschließenden Entwicklung von KI-Lösungen, die sich nahtlos in dein Unternehmen integrieren.

Möchtest du mit uns ins Gespräch kommen, um den Prozess und die erwarteten Ergebnisse eines KI-Projekts zu besprechen?