---
title: "Machine Learning"
description: "622 posts about Machine Learning, summarised, each linking to the original."
---

# Machine Learning
> 622 posts about Machine Learning, summarised, each linking to the original.

## Articles

### [How we improved translation experience with cost efficiency](https://yomu.fyi/post/how-we-improved-translation-experience-with-cost-efficiency.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Jie Zhang
- Published: Aug 5, 2024
- Type: Problem & solution

Grab faced translation quality and cost challenges in its booking chat system when tourism resumed after COVID restrictions eased. Device language settings proved unreliable for determining message language, prompting the team to use Lingua alongside booking context heuristics and third-party fallbacks for real-time detection. To replace expensive and inaccurate general translation APIs, Grab distilled a large open-source Hugging Face model into lightweight, language-specific models trained on ten million synthetic chat examples. A post-translation validation step verifies non-translatable entities such as numbers and emojis before falling back to external services if mutations occur. Caching layers were also introduced across translation paths to curb redundant on-the-fly execution.


### [Make Better Decisions by Embracing Uncertainty](https://yomu.fyi/post/make-better-decisions-by-embracing-uncertainty.md)
- Company: [Ramp](https://yomu.fyi/company/ramp.md)
- Author: Ryne Carbone
- Published: Jun 11, 2024
- Type: Problem & solution

Ramp uses multi-armed bandits to choose among third-party bank-linking providers when provider performance varies by financial institution and available data is noisy, incomplete, or stale. Instead of maintaining separate, static A/B tests, it applies Thompson Sampling, modeling each provider’s success probability with a Beta distribution and repeatedly sampling, routing traffic to the highest sampled value, then updating the model from the result. Simulations examined changing success-rate trends, noise, volume, rolling-window size, trend type, and initial conditions; the algorithm generally exploited clearly better providers, split traffic more evenly when rates were close, and recovered eventually from incorrect starting conditions. In a sequential quasi-experiment, Thompson Sampling increased financial-institution linking success by 10% and reduced customers with manually uploaded bank statements by 25%.


### [Efficient Visual Representation Learning And Evaluation](https://yomu.fyi/post/efficient-visual-representation-learning-and-evaluation.md)
- Company: [Etsy](https://yomu.fyi/company/etsy.md)
- Author: Eden Dolev
- Published: Jun 7, 2024
- Type: Explainer

Etsy uses image vector representations for visual search and similar-item recommendations, making low-latency, memory-aware models important for online query inference and downstream retrieval. The work compares EfficientNetB0, a compact CNN, with EfficientFormer-l3, which downsamples multiple blocks and applies attention only in its final stage; its representation uses the last block’s eight-head attention and average-pools the sequence. Fine-tuning adds an embedding and multiple classification heads, while multitask training supervises taxonomy, color, and review-photo labels; both models train on two 16GB P100 GPUs. Evaluation rebuilds a brute-force candidate index after each epoch and measures Recall@5 and Recall@10 on interaction and intra-item retrieval datasets, with early stopping available. EfficientFormer-l3 improves reported offline recall by about 5%, 17%, and 1.8% across three tasks, while experiments report lifts in CTR, purchase rate, click-recall@100, and clicks.


### [Enter eBay’s 2024 University Machine Learning Competition](https://yomu.fyi/post/enter-ebay-s-2024-university-machine-learning-competition.md)
- Company: [Ebay](https://yomu.fyi/company/ebay.md)
- Author: eBay News Team
- Published: May 6, 2024
- Type: Announcement

eBay’s 2024 University Machine Learning Competition invites U.S. college and university students to form teams and tackle a yearly ecommerce challenge, with top participants offered a shot at a 2025 summer internship at eBay. This year’s task focuses on eBay Motors: building a model that identifies which vehicles are compatible with particular parts and accessories. Participants must extract “fitment” information from listing titles, item specifics, and descriptions in the eBay Parts and Accessories categories, where noisy or extraneous data makes the task difficult. The competition is in its sixth year; last year, 1,439 students from 206 universities competed across 887 teams, and several winners later moved into full-time eBay roles.


### [NeurIPS 2023: Our Favorite Papers on LLMs, Statistical Learning, and More](https://yomu.fyi/post/neurips-2023-our-favorite-papers-on-llms-statistical-learning-and-more.md)
- Company: [Two Sigma](https://yomu.fyi/company/two-sigma.md)
- Author: Emily Majewski
- Published: Mar 21, 2024
- Type: Explainer

Researchers reviewed prominent machine learning papers presented at NeurIPS 2023 covering large language models and statistical learning theory. One investigation showed that claimed emergent abilities in models such as GPT-3 often result from nonlinear evaluation metrics rather than fundamental shifts in model capability. To reduce the computational burden of model adaptation, QLoRA enables 65-billion-parameter model fine-tuning on a single 48-gigabyte GPU via 4-bit NormalFloat quantization and paged optimization. Direct Preference Optimization eliminates complex reward modeling by casting reinforcement learning from human feedback into a preference classification task. Additional work resolved statistical anomalies like double descent using effective parameter counts and introduced stochastic gradient approximations for Gaussian processes.


### [Meet the Winners of the 5th eBay University Machine Learning Challenge](https://yomu.fyi/post/meet-the-winners-of-the-5th-ebay-university-machine-learning-challenge.md)
- Company: [Ebay](https://yomu.fyi/company/ebay.md)
- Author: eBay News Team
- Published: Feb 15, 2024
- Type: Announcement

The 2023 eBay University Machine Learning Challenge brought together 1,439 students from 206 universities in 887 teams to develop solutions for real-world ecommerce problems. The winning team, NullPointer, comprised five graduate students from universities across the United States and earned summer internships at eBay. Its project used Named Entity Recognition (NER) to identify, extract, and label named entities—including people, places, brands, and sizes—in a dataset of 10 million eBay Germany sneaker-category item titles. At eBay, NER is commonly used to automatically extract key information from listings and search queries. The event recorded its largest turnout yet, and students can register for the next competition via EvalAI in late spring.


### [Managing dynamic marketplace content at scale: Grab's approach to content moderation](https://yomu.fyi/post/managing-dynamic-marketplace-content-at-scale-grab-s-approach-to-conte.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Poonam Gambhire
- Published: Feb 1, 2024
- Type: Explainer

Grab manages content moderation across GrabFood and GrabMart, where merchants update over 100,000 item listings daily across eight countries. The platform must adhere to internal rules, local government regulations, and external platform guidelines across these markets. To handle scale and regional nuance, Grab uses an in-house automated system powered by algorithms and machine learning to scan daily listings for violations. Items flagged as requiring subjective judgment or cultural awareness are routed to human moderators for manual review. Additionally, Grab collaborates with Google to maintain compliance with Play Store policies.


### [What are these "GPUs" really?](https://yomu.fyi/post/what-are-these-gpus-really.md)
- Company: [Fly.io](https://yomu.fyi/company/fly-io.md)
- Published: Dec 11, 2023
- Type: Explainer

The post explains how graphics hardware evolved from CPU-offloaded video rendering into modern GPUs and why those devices now matter for AI/ML. It traces the path from the Commodore 64’s VIC-II and 3D accelerator cards through Quake, Tomb Raider, integrated graphics, and increasingly parallel hardware, while describing components such as framebuffers, command processors, shader hardware, memory, and display outputs. It then connects shader and CUDA cores to workloads including GPT-2, Stable Diffusion, and speech or text processing, and presents model chaining as more useful than treating models as isolated tools. A conference-video workflow combines ffmpeg, Whisper, subtitle segmentation, a large language model, Markdown, and human review to produce accessible written material. The conclusion distinguishes graphics-capable cards from A100 accelerators, which the post says cannot process graphics or encode video but are highly effective for AI/ML workloads.


### [Let AI Entertain You: Increasing User Engagement with Generative AI and Rejection Sampling](https://yomu.fyi/post/let-ai-entertain-you-increasing-user-engagement-with-generative-ai-and.md)
- Company: [Nextdoor](https://yomu.fyi/company/nextdoor.md)
- Author: Jaewon Yang
- Published: Oct 16, 2023
- Type: Problem & solution

Nextdoor implemented a generative AI framework to improve email notification subject lines for its New and Trending emails. Using off-the-shelf ChatGPT API outputs directly resulted in marketing-like phrases, hallucinations, and lower click-through rates compared to original user-generated post snippets. To resolve this, the team prompted the language model strictly to extract authentic post phrases without rewriting, eliminating hallucinations while preserving the original tone. Nextdoor then introduced a rejection sampling mechanism featuring a dedicated reward model that predicts user engagement and accepts generated subjects only when their score exceeds the user-written baseline. This combined system lifted sessions by 1%, boosted Weekly Active Users by 0.4%, and increased advertising revenue by 1%.


### [Scaling marketing for merchants with targeted and intelligent promos](https://yomu.fyi/post/scaling-marketing-for-merchants-with-targeted-and-intelligent-promos.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Sharon Teng
- Published: Oct 11, 2023
- Type: Problem & solution

Grab previously relied on globally assigned, heuristic promotional campaigns where all users could redeem offers until limits were reached. This lack of targeting and customisation failed to optimize promo spending or meet specific merchant business objectives. To solve this, the engineering team built Bullseye, an automated AI-driven promotional assignment system that customises and assigns offers to consumers. The architecture integrates an operations user interface, backend services, Amazon S3 storage, and Spark jobs executing a suite of predictive modules including eater segmentation, campaign impact simulation, and customer response modeling. Since deploying Bullseye in 2021, the system has increased food campaign sales while decreasing promo spend across food and Mart campaigns.


### [Stepping up marketing for advertisers: Scalable lookalike audience](https://yomu.fyi/post/stepping-up-marketing-for-advertisers-scalable-lookalike-audience.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: William Wu
- Published: Sep 22, 2023
- Type: Problem & solution

Grab's legacy lookalike audience generation platform suffered from long creation SLAs of two working days, high costs, and low weekly update frequencies. To resolve these bottlenecks, the engineering team designed an embedding-based platform powered by an in-memory retrieval service and automated update pipelines. The system creates audience representations by averaging constituent passenger embeddings, determining user membership through real-time cosine score thresholds. To eliminate feature store latency while fitting all embeddings into memory, a hash-based compression method cuts passenger embedding storage needs by roughly 90%. Consequently, audience availability dropped to within 15 minutes of campaign creation, audience generation costs fell by 98%, and ad impressions and clicks doubled.


### [How Ramp Accelerated Machine Learning Development to Simplify Finance](https://yomu.fyi/post/how-ramp-accelerated-machine-learning-development-to-simplify-finance.md)
- Company: [Ramp](https://yomu.fyi/company/ramp.md)
- Author: Peyton McCullough
- Published: Sep 15, 2023
- Type: Problem & solution

Ramp describes how its machine-learning organization used Metaflow to shorten the path from prototype to production for models spanning credit risk, fraud, growth, product, and Ramp Intelligence. Its initial riskiness model exposed long vendor-managed jobs, weak logging, limited Docker support, and substantial platform friction; the model took months to build. Metaflow, with developer-experience improvements, lets teams define flows in Python, run them locally or on AWS Batch, visualize results with cards, and connect production execution to Step Functions and Airflow. Ramp chose AWS-managed infrastructure, initially using Fargate before encountering startup and resource constraints, and later built a MetaflowOperator to simplify Airflow triggering and log access. After adoption, Ramp shipped eight additional models in ten months, recorded more than 6,000 Flow runs, and reports that data scientists can largely self-service while platform engineers spend less time debugging infrastructure.


### [From Pre-trained to Fine-tuned: Nextdoor’s Path to Effective Embedding Applications](https://yomu.fyi/post/from-pre-trained-to-fine-tuned-nextdoor-s-path-to-effective-embedding.md)
- Company: [Nextdoor](https://yomu.fyi/company/nextdoor.md)
- Author: Karthik Jayasurya
- Published: Sep 7, 2023
- Type: Problem & solution

Nextdoor transitioned its ranking and recommendation pipelines from traditional continuous and discrete interaction features to transformer-based representation learning. The engineering team deployed pre-trained Sentence-BERT models to generate multilingual post and comment representations, which were aggregated daily by interaction type to form user embeddings. To improve search recall, the team fine-tuned sentence transformers on unlabeled query session logs using contrastive learning and integrated HNSWlib for approximate nearest neighbor retrieval. Subsequent iterations incorporated labeled feedback, BERTopic for coarse personalization, and experiments with CLIP image embeddings. Infrastructure scaling challenges were addressed by performing embedding transformations directly within FeatureStore and optimizing feature payload formats to minimize microservice network bandwidth.


### [The So-fine Real-time ML Paradigm](https://yomu.fyi/post/the-so-fine-real-time-ml-paradigm.md)
- Company: [Etsy](https://yomu.fyi/company/etsy.md)
- Author: Kyle Gallatin
- Published: Sep 5, 2023
- Type: Problem & solution

During Etsy’s CodeMosaic hackathon, a team investigated stateful training and online learning to reduce the roughly 40-hour delay between user actions and model-weight updates. They split the work across real-time training data, an incremental-learning service, and evaluation, using Etsy’s Beacon Main Kafka stream, Kafka SQL, Rivulet, and TensorFlow-based Ads models. The data team joined multiple sources despite Avro schema difficulties, while the service team loaded an older model and nearly implemented incremental weight updates; the planned head-to-head evaluation against daily batch training was not completed. The project estimated that replacing daily cold-start training for four ads models could save about $212,000 annually in Google Cloud costs, with possible metric gains from earlier events. Production would still require infrastructure and safeguards against sudden performance degradation.


### [pgvector: Fewer dimensions are better](https://yomu.fyi/post/pgvector-fewer-dimensions-are-better.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: gregnr,oli\_rice,egor\_romanov
- Published: Aug 3, 2023
- Type: Benchmark

Embeddings support semantic search, retrieval augmented generation, and clustering in pgvector, but high-dimensional vectors increase storage, memory, and distance-computation costs as datasets grow. The post explains how IVF indexes reduce the records examined and compares embedding models using Massive Text Embedding Benchmark results, noting little correlation between dimension count and performance. In a 224,482-vector Wikipedia dataset, all-MiniLM-L6-v2 with 384 dimensions delivered 78% higher queries per second than OpenAI text-embedding-ada-002 with 1536 dimensions at accuracy@10 of 0.99, while using 4GB rather than 7.5GB of RAM. The 384-dimensional gte-small model produced more than a 200% queries-per-second improvement over all-MiniLM-L6-v2 at the same accuracy, though model choice still depends on task performance, sequence length, language support, and operational costs.


### [Unsupervised graph anomaly detection - Catching new fraudulent behaviours](https://yomu.fyi/post/unsupervised-graph-anomaly-detection-catching-new-fraudulent-behaviour.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Rizal Fathony
- Published: Aug 2, 2023
- Type: Problem & solution

Fraud detection systems relying on historical training data often struggle to detect new fraudulent modus operandi because emerging patterns lack labeled supervision. Grab developed GraphBEAN, an unsupervised autoencoder model designed to detect anomalous patterns across bipartite interaction graphs representing consumers and merchants. The architecture uses graph convolution layers to encode node and edge attributes, then reconstructs features and edge existence through separate feature and structure decoders. High reconstruction errors flag rare, anomalous behaviors at both the node and edge levels. An automated pipeline applies heuristic fraud tags to these scores and routes them to human analysts and automated mitigation systems.


### [pgvector 0.4.0 performance](https://yomu.fyi/post/pgvector-0-4-0-performance.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: egor\_romanov,pavel
- Published: Jul 13, 2023
- Type: Benchmark

The post benchmarks pgvector 0.4.0 for approximate nearest-neighbor search and assesses whether PostgreSQL can support production-scale vector workloads. Using the ANN Benchmarks methodology, pre-warmed RAM, the dbpedia-entities-openai-1M dataset of 1 million 1,536-dimensional OpenAI embeddings, and Supabase ARM instances, it varies probes, lists, concurrency, and compute size. On an 8-core, 32 GB instance, pgvector reached 0.91 accuracy@10 at 380 QPS with 10 probes and 0.98 accuracy at 140 QPS with 40 probes; a 64-core, 256 GB server reached about 1,800 QPS at 0.91 accuracy. The results indicate predictable scaling, while higher lists can improve query speed at the cost of index-build time, and the roadmap identifies HNSW, product quantization, and parallel index builds as planned improvements.


### [PII masking for privacy-grade machine learning](https://yomu.fyi/post/pii-masking-for-privacy-grade-machine-learning.md)
- Company: [Grab](https://yomu.fyi/company/grab.md)
- Author: Fabrice Harbulot
- Published: Jun 1, 2023
- Type: Problem & solution

Data engineers at Grab require real-world streaming data to tune machine learning models, but user personal data cannot be exposed in non-production environments. To solve this, the data streaming team implemented an automated masking pipeline between production and staging Kafka clusters. Developers explicitly tag Personally Identifiable Information (PII) types in Protocol Buffers schemas, which a CI script validates against a keyword list before packaging the schemas into Scala JARs. An in-house Apache Flink application running in production consumes the unmasked streams, applies dynamic and consistent transformations like keyed HMACs or seeded random values, and produces sanitised records to staging. Untagged new fields in schema updates are automatically dropped by the deserialiser until the masking application is redeployed, preventing accidental data leaks.


### [Flutter Hackathon Winners](https://yomu.fyi/post/flutter-hackathon-winners.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: tyler\_shukert
- Published: May 29, 2023
- Type: Announcement

The Flutter Hackathon winners were announced across categories including Best Overall Project, Best FlutterFlow Project, Best Dart Edge Project, Most Technically Impressive, and Most Visually Pleasing. Syno won Best Overall for concise video summaries, while Study Genius was runner-up; other recognized projects included Lord Of The Memes, ChatGlobe, Accident Detection, Caffeio, and their respective runners-up. ChatGlobe translates messages into multiple languages as they are sent using Dart Edge and the OpenAI API, while Accident Detection uses machine-learning techniques with accelerometer and GPS data to analyze accident-related changes in speed, direction, and impact force. Winners and runner-up teams receive limited Flutter Hackathon swag, and Syno’s creator also receives a Flutter-themed Supabase keyboard; the announcement also points to Flutter, Supabase, FlutterFlow, Edge Functions, and authentication guides for future projects.


### [Launch Week 7 Hackathon Winners](https://yomu.fyi/post/launch-week-7-hackathon-winners.md)
- Company: [Supabase](https://yomu.fyi/company/supabase.md)
- Author: tyler\_shukert
- Published: Apr 24, 2023
- Type: Announcement

Supabase Launch Week 7 announces the winners of its hackathon, which introduced an AI theme and recognized projects across five categories. Page Assist won Best Overall for an open-source Chrome extension that lets users chat with webpages through the ChatGPT API, while other winners included Rect, Supathreads, Replay.ai, Supa0SQL, and Generation: Hotdog. The projects apply AI to educational content, blog-to-thread conversion, browsing-history search with cosine similarity, natural-language chart creation, and hotdog-image generation using OpenAI and Stable Diffusion APIs; runner-ups included tools for video chat, drum patterns, tweet scheduling, and music guessing. Members of the Best Overall winning team receive a limited-edition Launch Week 7 Keyboard, while winners and runner-ups in other categories receive Launch Week 7 swag kits, alongside guides for getting started with Supabase resources.


[Newer posts](https://yomu.fyi/topic/machine-learning/page/23.md) · [Older posts](https://yomu.fyi/topic/machine-learning/page/25.md)