AI research and projects
The research, projects, and explanations behind the people. Read the source, inspect verified code and models, and find a collaborator.
A source check confirms that a source supports the summary. It is not an independent replication or endorsement. Our methodology ↗
Leakage and the Reproducibility Crisis in Machine-Learning-Based Science
Dr. Sayash Kapoor · Princeton UniversityPublished a comprehensive meta-analysis of machine learning scientific literature discovering data leakage across 329 papers spanning 17 fields, establishing formal taxonomy and verification criteria to prevent train-test contamination.
Meta-analyses depend on publicly available replication code; studies with proprietary or non-shared datasets could not be independently tested.
Context: Reproducibility audit across computer vision, clinical healthcare predictions, and natural language processing.
View missionMistral 7B: Grouped-Query and Sliding Window Attention Foundation Model
Arthur Mensch · Mistral AIDesigned and published Mistral 7B, an open-weights autoregressive transformer utilizing Grouped-query Attention (GQA) and Sliding Window Attention (SWA) to achieve faster inference and handle 8k context sequences while beating Llama 2 13B across all standard benchmarks.
Requires high cache memory bandwidth on commodity GPUs; sliding window attention drops exact attention history beyond context horizon without caching adaptations.
Context: 7.3B parameters, 32 layers, hidden dimension 4096, 32 attention heads, 8 key-value heads (GQA), sliding window attention of 4096 tokens.
View missionMixtral of Experts: Sparse Mixture-of-Experts Open Weight Model
Arthur Mensch · Mistral AIPublished the architectural blueprint for Mixtral 8x7B, a sparse mixture-of-experts model where a router chooses 2 out of 8 feedforward experts per token, activating only 13B parameters out of 47B during forward pass and outperforming Llama 2 70B with 6x faster throughput.
High RAM/VRAM memory footprint for hosting all 47B parameters despite activating only 13B during computation; requires specialized multi-GPU tensor or pipeline parallelism.
Context: 46.7B total parameters, 12.9B active parameters per token, 32k token context window, Top-2 softmax gating router.
View missionMLflow: A Machine Learning Lifecycle and Foundation Model Evaluation Platform
Dr. Matei Zaharia · Databricks / UC BerkeleyCreated MLflow, an open-source platform managing the end-to-end machine learning lifecycle, including experiment tracking, model registry, artifact packaging, and evaluation metrics for classical ML and modern LLM pipelines.
Self-hosted deployments require configuring and maintaining persistent artifact object storage (S3/GCS) and relational backend metadata databases.
Context: Python, REST, Docker packaging, native support for PyTorch, Hugging Face, LangChain, and OpenAI SDKs.
View missionModel Cards for Model Reporting: Standardizing ML Evaluation Disclosures
Dr. Margaret Mitchell · Hugging FaceAuthored the seminal ACM FAccT paper introducing 'Model Cards for Model Reporting', an operational documentation standard that details model architecture, intended tasks, training parameters, out-of-scope uses, and disaggregated demographic evaluations.
Static documentation can diverge from deployed weights if model inputs undergo runtime changes or unlogged adapter swaps.
Context: Disaggregated evaluation metrics across intersecting demographic slices and environmental contexts.
View missionMulti-Agent Workflow Orchestration for Financial Document Processing
Helena García Borrella · The Interactive Studio / Codebridge TechDesigned and deployed an autonomous multi-agent pipeline parsing heterogeneous multilingual financial invoices, extracting line items into structured JSON, and verifying tax calculations with zero human intervention.
Invoices with non-standard European VAT layouts require automated human-in-the-loop review queues.
Context: GPT-4o-mini, LangGraph, Pydantic validation schemas, OCR vision models.
View missionNeedle In A Haystack: Pressure Testing Long Context Retrieval in LLMs
Greg Kamradt · Data IndependentCreated the industry-standard Needle In A Haystack benchmark that places distinct target facts across varying depths (0% to 100%) and context lengths (1k to 1M+ tokens) to measure retrieval degradation in frontier foundation models.
Single-needle retrieval does not test complex multi-hop reasoning or distributed cross-document synthesis across long contexts.
Context: Evaluated across GPT-4-Turbo, Claude 2.1/3, Gemini 1.5 Pro, Llama-3-70B.
View missionOn the Dangers of Stochastic Parrots: Can Language Models Be Too Big?
Dr. Timnit Gebru · Distributed AI Research Institute (DAIR)Published the landmark ACM FAccT paper critically examining the environmental costs, training data biases, lack of semantic grounding, and illusion of intent when scaling large autoregressive language models without curated provenance.
Focusing on systemic risks does not quantify operational benefits in domain-specific tasks where automated verification pipelines exist.
Context: Analysis of Web-scraped pre-training corpora (Common Crawl, C4) and environmental compute costs.
View missionPerplexity Sonar and Real-Time Agentic Search Pipeline
Aravind Srinivas · Perplexity AIArchitected Perplexity's conversational search engine and Sonar API, synthesizing real-time multi-source web crawling, dense vector retrieval, dynamic citation parsing, and fine-tuned LLM response generation with strict anti-hallucination attribution.
Subject to latency variance across heterogeneous external web APIs; adversarial or SEO-manipulated web sources require continuous algorithmic filtering.
Context: Llama-3-Sonar fine-tuned checkpoints, TensorRT-LLM, sub-second latency SLA, dynamic search grounding over live web indexes.
View missionPythia: A Suite for Analyzing Large Language Models Across Training and Scaling
Stella Biderman · EleutherAI / Booz Allen HamiltonDirected the development and release of the Pythia model suite: 16 LLMs ranging from 70M to 12B parameters, all trained on identical data ordering with 154 intermediate checkpoints public for scientific analysis of memorization and bias.
Pre-training dataset contained early web corpora reflecting historical internet text distributions prior to modern RLHF curation.
Context: GPT-NeoX distributed training framework, FlashAttention, 300B tokens of the Pile dataset.
View missionReal-Time Computer Vision Quality Control Pipeline for Manufacturing
Rahul Patel · Space-O Technologies / ManageMyClawDeployed an edge computer vision defect detection model running on industrial cameras that inspects automotive components on conveyor belts at 60 FPS with 99.4% precision.
Drastic variations in factory floor lighting conditions require automated photometric normalization algorithms.
Context: YOLOv8 on NVIDIA Jetson Orin Nano, TensorRT export, RTSP video streams.
View missionRelative Representations Enable Non-Degenerate Latent Space Alignment
Dr. Been Kim · Google DeepMindCo-authored breakthrough research demonstrating that representations can be compared across disparate neural networks by computing pairwise angles and similarities to anchor points, proving latent geometries are invariant across architectures.
Choice and distribution of anchor points can introduce variance in representation reconstruction accuracy across domain shifts.
Context: Cross-model embedding alignment across Vision Transformers, CNNs, and language models without weight access.
View missionServerless Function Calling Middleware for Conversational Commerce
Raúl Pineda Mejía · Codebridge Tech / The Interactive StudioDeveloped a serverless middleware connecting natural language user intents to live inventory databases, executing secure database lookups and payment checkouts via strict JSON schema validation.
Requires deterministic database transaction rollbacks if client connection drops during checkout execution.
Context: AWS Lambda, Claude 3.5 Sonnet Tool Use API, PostgreSQL RDS.
View missionServerless High-Throughput vLLM Inference Container Orchestration on Modal
Charles Frye · Modal LabsEngineered containerized serverless LLM deployment architectures on Modal, orchestrating multi-GPU vLLM inference with cold-start mitigation, dynamic batching, and automated autoscaling from 0 to 100 NVIDIA A100/H100 instances.
Cold-start container image pulling and model weight downloading require warm snapshot caching techniques to maintain sub-second response readiness.
Context: Modal Python SDK, CUDA 12, PagedAttention, Tensor Parallelism across NVLink clusters.
View missionSocial Chemistry 101: Learning to Reason about Social and Moral Norms
Dr. Yejin Choi · University of Washington / Allen Institute for AIIntroduced Social Chemistry 101, a computational framework and dataset containing 292,000 rules-of-thumb covering social conventions, moral rules, and situational ethical trade-offs, enabling LLMs to reason explicitly about multi-cultural norms.
Moral norms vary across geographic cultures, religious traditions, and temporal epochs; models require adaptive normative context parameters.
Context: Structured social rule-of-thumb generation with attribute classifications (morality, social judgment, agreement level).
View missionThe Full Stack LLM Bootcamp: End-to-End Enterprise LLM Production Engineering
Charles Frye · Modal LabsDesigned and delivered the comprehensive Full Stack LLM Bootcamp, educating tens of thousands of software engineers on prompt design, fine-tuning with LoRA/QLoRA, vector retrieval, evaluation metrics, and cost-efficient deployment.
Rapid evolution in model capabilities requires constant curriculum revisions to address frontier tool-calling interfaces.
Context: PyTorch, Hugging Face, vLLM, LangChain, Modal serverless GPU runtimes.
View mission