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 ↗
Accenture Foundation Model Adoption & Value Realization Framework
Lan Guan · AccentureDeveloped Accenture's proprietary enterprise AI reinvention methodology, deploying foundation models across supply chain, customer service, and software engineering for Fortune 500 clients while benchmarking return on AI investment (ROAI).
Large-scale enterprise legacy IT modernization timelines often outlast fast-moving generative foundation model upgrade cycles.
Context: Enterprise LLM operationalization, multi-cloud foundation model architectures, ROAI metrics.
View missionDatabricks Mosaic AI & Lakehouse Architecture for Foundation Models
Ali Ghodsi · DatabricksLed the engineering integration of MosaicML training and fine-tuning pipelines directly into the Databricks Lakehouse, enabling organizations to pretrain and align custom open-weight models securely on internal enterprise data lakes.
On-premise and private cloud cluster provisioning constraints can bottleneck distributed tensor parallelism across non-homogeneous GPU topologies.
Context: DBRX, MPT-7B/30B, MosaicML Composer, Delta Lake, Vector Search.
View missionDeterministic Agentic Execution Architecture for Enterprise Customer Systems
Bret Taylor · Sierra / OpenAIArchitected Sierra's Agent OS, combining generative LLM natural-language comprehension with deterministic state machines and strict transactional API sandboxes to prevent agent hallucination in consumer-facing enterprise tasks.
Enforcing strict deterministic guardrails constrains open-ended agent creativity and requires substantial domain-specific schema modeling.
Context: Multi-LLM routing, deterministic state automata, transactional CRM/ERP webhooks.
View missionEnterprise Knowledge Graph Grounding for Multi-Source RAG Systems
Arvind Jain · GleanArchitected Glean's enterprise intelligence engine, coupling deep permissions-aware crawling of 100+ SaaS applications with dense vector embeddings and a dynamic enterprise knowledge graph for sub-second grounded generative question answering.
Maintaining real-time permission sync across complex Active Directory, Okta, and SaaS permission structures introduces significant distributed indexing overhead.
Context: Hybrid sparse-dense retrieval, enterprise permissions ACL enforcement, LLM synthesis.
View missionLamini Memory Tuning: Precise Fact Recall Without Hallucination
Sharon Zhou · Lamini / ex-StanfordDeveloped Lamini Memory Tuning, a technique embedding millions of precise factual enterprise associations into LLM weights with zero hallucination while preserving generalized instruction-following capability.
Aggressive factual weight baking risks model overfitting and can degrade out-of-distribution reasoning on non-memorized prompts.
Context: LoRA/PEFT adaptation, cross-entropy weight tuning, hallucination evaluation harnesses.
View missionNVIDIA Blackwell Platform: Accelerated Architecture for Trillion-Parameter Generative AI
Jensen Huang · NVIDIAArchitected the Blackwell B200 GPU and NVLink 5 switch chip infrastructure, introducing second-generation Transformer Engine FP4/FP8 micro-tensor scaling and high-density liquid-cooled rack computing designed to reduce trillion-parameter LLM inference costs and energy consumption by up to 25x.
High thermal design power (up to 120kW per NVL72 rack) requires specialized data center liquid-cooling retrofits and high-density power distribution.
Context: NVIDIA Blackwell B200, NVLink 5 (1.8TB/s bidirectional interconnect), FP4/FP8 Transformer Engine, TensorRT-LLM.
View missionOpenAI o1 & Test-Time Compute Scaling Architecture
Sam Altman · OpenAIDirected the commercialization and architectural rollout of OpenAI o1, shifting foundation model capability scaling from purely pretraining tokens to inference-time chain-of-thought reinforcement learning reasoning loops.
Increased inference latency (tens of seconds per complex query) and higher token costs make pure test-time reasoning impractical for real-time conversational streaming without hybrid routing.
Context: OpenAI o1, chain-of-thought reasoning, reinforcement learning from human and rule-based feedback (RLHF/RLAIF).
View missionRedPajama: Open Pretraining Datasets for Reproducible Foundation Models
Vipul Ved Prakash · Together AICo-led the open-source release of RedPajama, a 1.2-trillion and 30-trillion token pretraining dataset replicating the LLaMA pretraining corpus to enable transparent, fully reproducible foundation model research.
Web-scale data deduplication and filtering at 30T scale require petabyte-scale distributed compute and risk residual benchmark contamination.
Context: 30T token web corpus, deduplication pipelines, MinHash LSH, toxicity filtering.
View missionSalesforce Agentforce: Autonomous Enterprise Multi-Agent Orchestration
Clara Shih · Meta / ex-SalesforceSpearheaded the product architecture of Agentforce, enabling enterprises to deploy autonomous AI agents with unified CRM data grounding, Atlas reasoning engines, and policy guardrails that operate without rigid hardcoded decision trees.
Enterprise agent reliability requires high data cleanliness and granular field-level permissioning across fragmented legacy enterprise systems.
Context: Atlas Reasoning Engine, Einstein Trust Layer, CRM Data Cloud grounding.
View missionSnorkel: Rapid Training Data Creation with Weak Supervision
Alex Ratner · Snorkel AI / University of WashingtonCreated the Snorkel open-source system and paradigm, using generative probabilistic modeling to combine noisy heuristic rules, knowledge bases, and small models into high-accuracy probabilistic training labels without manual annotation.
Heuristic labeling functions can introduce correlated bias if domain experts write overlapping or conditionally dependent heuristics without proper graph independence modeling.
Context: Probabilistic generative label models, matrix completion, weak supervision algorithms.
View missionTogether Inference Engine & FlashAttention-Optimized Model Serving
Vipul Ved Prakash · Together AIEngineered Together AI's custom inference kernel stack, implementing speculative decoding, custom FP8 FlashAttention kernels, and continuous batching to deliver the industry's fastest serving latency for Llama 3, DeepSeek, and Mixtral.
Highly optimized custom kernels require aggressive quantization and continuous hardware profiling for non-standard GPU architectures.
Context: Llama 3/3.1, Mixtral 8x22B, custom CUDA/Triton kernels, speculative sampling, continuous batching.
View missionAgent.ai: Multi-Agent Professional Workflow Directory & Execution Engine
Dharmesh Shah · HubSpot / Agent.aiArchitected a global discovery and execution platform for specialized AI agents that orchestrate multi-step research, prospect intelligence, competitive synthesis, and automated business workflows.
Autonomous agent execution requires strict human-in-the-loop review for state-changing enterprise actions to prevent cascading hallucinated writes.
Context: Distributed agentic orchestrator, LLM tool-calling APIs, dynamic web sandboxes.
View missionChatSpot.ai: Natural Language Conversational CRM Layer
Dharmesh Shah · HubSpot / Agent.aiDeveloped one of the first production enterprise generative AI integrations combining OpenAI GPT-4 with structured enterprise CRM relational databases, executing natural language SQL queries and automated prospecting.
Schema complexity across enterprise custom objects necessitates strict few-shot prompt guards to avoid incorrect relational joins.
Context: GPT-4, HubSpot API, semantic intent parsers, relational data synthesizers.
View missionDiffusers: State-of-the-Art Diffusion Models in PyTorch
Suraj Patil · Hugging FaceEngineered core schedulers (DDIM, DPMSolver, Euler), memory-efficient cross-attention kernels, and quantized pipeline architectures powering Stable Diffusion and Flux ecosystem inference.
Diffusion sampling requires multiple sequential denoising steps, bottlenecking latency relative to single-pass adversarial models.
Context: PyTorch, CUDA, xFormers, FlashAttention, ONNX runtime.
View missionDiffusion Model Distillation & Latent Consistency Optimization
Sayak Paul · Hugging FaceCo-authored and integrated Latent Consistency Model (LCM) LoRA adapters into Hugging Face Diffusers, compressing multi-step diffusion sampling into 2 to 4 inference steps for real-time generation.
Aggressive 2-step distillation can introduce subtle high-frequency smoothing in photorealistic background details.
Context: PyTorch, CUDA, LCM-LoRA, SDXL, Apple Silicon Metal.
View missionGitHub Copilot: Real-Time Inline Generative Code Completion Architecture
Alex Graveley · Minion AI / Former GitHubConceived, architected, and built the initial GitHub Copilot production system, engineering client-side AST context window chunking, token speculative decoding, and low-latency streaming to VS Code.
Context window constraints in early versions restricted completions to single-file buffers before repository-wide semantic indexers were introduced.
Context: OpenAI Codex, TypeScript, VS Code language client, telemetry evaluation loops.
View mission