Community Development Services Provider

Data Science

• Utilized Google OCR and created parser functions to retrieve data on pages of various non-standard formats of construction plans in PDF format
• Implemented image pre- and post-processing techniques for improving OCR results

Backend

• Modified existing Django backend and REST API for improving API responses

DevOps

• Monitored Kubernetes clusters for performance metrics of parser functions
• Setup AWS SES for early warning notification based on network traffic

Montessori Micro-Schools Company

Data Science

• Real-time pose estimation with Alphapose on CCTV footages
• Hosted pose estimation model on AWS SageMaker

Private Holding Company

Data Science

• Trained models to deliver account hierarchies in finances

Backend

• Utilized AWS services such as API Gateway, Lambda, and S3 for exposing trained models in a single API

Mining Company

Data Science

• Employed existing plate tectonic models to generate proxmity maps to possible locations of specific mineral deposits based on mantle plume location
• Paleospace analysis by rotating present-day coordinates to their birth coordinates

Fashion Company

Generative AI

• Trained text embeddings using image-text pairs from scraped image dataset and output of style classification models
• Generated modified images (inpainting using mask) to adhere to specific style aesthetics using text prompts and trained text embeddings
• Visualized analysis of trained embeddings for most similar word embeddings exploration
• Created customized generative pipeline for guiding generated images to style aesthetics from another text and image encoder model

Computer Vision

• Style and image embeddings clustering and analysis with multimodal embedding model, capturing similarities and centralities

NGO

Image Generation Pipelines

• Implemented image generation pipelines intended for Agentic Tool use.
  • v1 implementation: Self-hosted Flux.1-dev, integrating InstantX's IP Adapters for face identity transfer and LoRAs for amateurization. Anchor images per persona profile are generated first and reused for future generations, preserving facial identity. Model prompts are auto-generated through a prompt-generating agent, instructed to convert a simple scene description into prompts designed for the model and matches the persona profile.
  • v2 implementation: Migrated to use Wan2.7-Image-Pro as the base image generation model. A stronger face-identity retention is seen in internal validations, with the model's built-in understanding of using seed images as context. Prompt-generating agents are still utilized but tuned towards how the new base model adheres to prompts.

Scout Agents

• Created a platform to dispatch social media scout agents.
  • Makes use of existing playwright-based social media crawlers using browser automation as tools.
  • A campaign goal turns into a task graph which is sent over to an available scout agent for completion. Tasks include topic research and social media profile resolution (both via web search), profile/post/reel scraping, and analysis.
  • Audience Filtering: scraped comments can be filtered by criteria (match a specific demographic by looking at profile avatars and comment content) along with location matching via profile bio or using existing geolocation API for inferring locations based on photos posted by the users.
  • Follow users: Profiles matched using the Audience-Filtering Agent can be followed also using existing playwright-based browser automation
  • Agents: Task Graph Planner Agent, Scout Agent, Audience-Filtering Agent, Follower Agent, Analyzer/Synthesizer Agent
  • Hybrid LLM setup using Alibaba Cloud enterprise models thru their APIs for research/analyze modes and a self-hosted model for crawling modes.

Dental Clinic Services Company

Computer Vision

• Improved a generative model for teeth generation as preview of treatment results

MLOps

• Implemented inference pipeline and deployed the model to AWS SageMaker for scalable hosting
• Implemented GitLab CI/CD pipelines for model and API deployment automations
• Implemented Terraform scripts for API deployment

Fashion Company

MLOps

• Implemented asynchronous model inference, GPU instance autoscaling, and deployment scripts to AWS SageMaker
• Utilized AWS API Gateway for exposing versions of the model thru an API