About

Last Frontier Technologies

We stand apart from the usual consultants. Having extensively navigated the industry’s complexities, our aim is to help you avoid pitfalls by imparting the insights we’ve gained, ensuring you make informed investments about where to allocate resources and what to develop. And the kicker is that we legitimately love doing this. We love solving problems for people.

Problems We Solve

Our experience is in applied Machine Learning at all levels of complexity. We’ve worked on increasing LLM reasoning capabilities with MCTS, we’ve worked on vision signals for industrial manufacturing, and we’ve worked with on classic machine learning problems like predicting player faceoff wins for the NHL.

Consider hiring Last Frontier Technologies if:

  • You don’t have the knowlege/capabilities to add ML complexity to your product
  • You tried a promising idea that didn’t perform as well as advertised
  • You know what would improve your product, but aren’t sure on how to go about it

Pricing

  • A typical project ranges from $50k to $150k over a minimum 3-month period

Notable Projects

Historical Experience

Below are projects completed during our consulting careers:

SLAM/3D Reconstruction

  • NFL Next Gen Stats: Player tracking and identification using CenterTrak with PyTorch for tracking unlabeled players; 3D field reconstruction via camera homography/SLAM methods mainly utilizing colmap for alternate angle play visualization
  • NFL Player Health and Safety: Ground plane estimation using MeTRAbs pose estimator for understanding location of on-field injury

Tracking and Identification

  • Amazon Fresh: Scalable C++ based multi-camera 3D tracking system for understanding sequences of events such as entry/exit, hand-washing, stateful understanding of worker actions in Amazon Fresh warehouses
  • JBS: Visual tracking system designed for pigs deployed to JBS’ facilities using Siamese Networks
  • Pepsi: Creating a POC of an automated vending machine that operates similarly to AmazonGo; gesture recognition, tracking, and weight analysis
  • McDonald’s: Designed Master Chef, a computer vision state-machine involving object/employee tracking/detection, and inventory management of the back-of-kitchen. Much of this was done through the generous collaboration from our AmazonGo colleagues

Pose Estimation

  • NFL Player Health and Safety cont: PoEM (Human POse EMbedding) View Invariant Probabilistic Embeddings for Human Pose. Usage: building visual dataset of player injury and querying similar injuries across history for analysis of injury cause

Object Detection and Segmentation

  • Wilson Sporting Goods: Object detection and recommender API to suggest alternative Wilson products to online shoppers
  • Honda: Created the ‘Inspectron’ library, a rules-based visual QA framework utilizing SIFT, SuperGlue, Lookout For Vision to detect irregularities in physical engine parts
  • Merck: Vile seal detection system using segmentation for a transportable Vaccine Pod meant to be deployed to conflict zones
  • Seaboard: Hand-written tattoo recognition system for detecting tattoo-ids on pigs inside Seaboard’s ingestion plant. Work published in the 2021 Amazon Machine Learning Conference
  • Caesar’s Palace: Stack-detection system designed to count player chips during blackjack games using OpenCV and segmentation models
  • Coca-Cola: Carbonation analysis using semantic segmentation and FFT for soda freshness project; camera monitors foam height and signal is analyzed to understand soda machine part failure
  • General Electric: Serial number detection system using ShuffleNet YOLOv5 deployed to edge devices for automating part placement
  • Shell, ExxonMobil, Enbrige: Oil spill detection for pure Computer Vision for each gas station by retrofitting CCTV cameras

Synthetic Data

  • Tyson: Building and deploying a novel ShuffleNet based object detection system to an edge device as well as a Synthetic Data creation system utilizing 3D scanning and Blender that helped propel SageMaker Ground Truth Synthetics resulting in a 9500x reduction in research costs-blog post here

Image Similarity/Recommendations

  • Anheuser-Busch, Argentine Government, Seaboard, GE, and others: Built an MLOps/Serving framework for detection and massive image-similarity queries using FAISS, YOLOv3, and ResNet that runs on pure CPU and in returns billion image queries at millisecond time

Anomaly Detection

  • Liberty Mutual, Allstate: Driver tracking and signal analysis for each insurance provider using XGBoost
  • Occidental Petroleum: Created ‘True Lies’, a novel motif/time-series detection system a utilizing MASS/Matrix Profiles and Dynamic Time Warping for ‘period of time’ similarity/embeddings for anomaly detection and labeling time series datasets

Language Modeling

  • Eli Lilly: Adverse text event detection system utilizing BERT embeddings
  • Polaris: Customer response classification system for ranking severity of car problems

MLOps and Infrastructure

  • International, Flavors and Fragrances, Schlumberger, Axalta: MLOps scoping for designing a research stack for each company’s scientist groups

Workshopping

  • JP Morgan, Goldman Sachs: SageMaker research environment deployed to each bank’s quant research team
  • Monsanto, Motorola, Nokia, Vodafone: IoT and signal-analysis workshop to educate researchers on the benefits of an AWS stack

Get in Touch

Ready to leverage our expertise for your next project? Contact us today to discuss how we can help you achieve your machine learning goals.