365UI builds private AI systems for serious enterprise workflows

Turn scattered company knowledge into agents people can trust.

Superagent is the flagship 365UI platform: a self-hostable RAG and agent system already deployed and operated inside an S&P 500 enterprise environment for teams that need accurate answers, auditable workflows, and private AI operations over real business data.

Design a pilot See the platform

S&P 500

enterprise production deployment

98.9%

enterprise benchmark accuracy

137s → 6.6s

retrieval pipeline latency

Superagent runtimeprivate deploy

Answer path

Ingest → Retrieve → Rerank → Act

Evidence-grounded answers with traceable retrieval and replaceable infrastructure.

Knowledge sourcesPDF / Web / DB / Email

RetrievalVector + BM25 + Rerank

APIOpenAI compatible

ControlsTrace / eval / review

RAGAgentsMemory

Platform

One runtime, multiple products.

The public story stays simple: private AI infrastructure for teams. The underlying work spans RAG, memory, inbox automation, realtime voice, agent tools, and hardware operations.

Enterprise RAG + Agent Runtime

Superagent

A private AI layer for teams that need precise answers from messy documents, product data, tickets, databases, and internal workflows.

Hybrid searchRerankingOpenAI-compatible APIAgent tools

Organizational Memory

OrgMem

An AI-era organizational memory layer where humans and agents co-write knowledge, Markdown + Git is the source of truth, and validated execution memory evolves into durable know-how.

Markdown + GitMCP/Hooks5-Plane MemoryPromotion

Executive Workflow Automation

Email Agent

A private assistant for high-volume inboxes: daily digest, smart triage, attachment parsing, semantic email search, and draft replies.

Daily digestPriority routingAttachment RAGHuman approval

Realtime Personal Interface

Voice Agent

A hands-free voice layer for reminders, local device signals, and realtime conversation with tool execution.

Wake wordRealtime audioRemindersDevice context

Autonomous Workflow Automation

Process Agents

Agents that turn repetitive expert workflows into reviewed automation: code-fix runs, recruiting shortlists, report generation, and tool-backed operations.

Static analysisATS workflowsHuman reviewAudit trail

Model Evaluation + GPU Inference

AI Infrastructure

Benchmarking and deployment patterns for Llama, DeepSeek, Qwen, GLM, Kimi, embedding, and reranker models across modern GPU clusters.

40+ modelsvLLMSGLangllama.cpp

AI Collaboration Governance

AI Collab Standard

A documentation governance standard for CLAUDE.md, AGENTS.md, .cursorrules, lesson_learned, and ADRs so human and AI teams can collaborate without context decay.

Layered rulesADR templatesLessonsOne-command init

Proof of Work

A portfolio of production agents, not isolated demos.

365UI is built from systems that have been deployed and operated in production, including S&P 500 enterprise copilot infrastructure, high-scale retrieval, autonomous process agents, model-serving infrastructure, messaging assistants, and edge voice products.

Private enterprise copilot

Enterprise Universal AI Agent Platform

Architected a fully private, self-hosted AI assistant from zero to production. The platform handles web and SharePoint-style crawling, complex PDF/Office parsing, tenant-specific configuration, and zero-code deployment across industries.

Zero to productionMulti-tenant configCompany-wide copilot path

Search, tools, and reasoning

Intelligent Retrieval & Agent Orchestration

Built hybrid retrieval over 1M+ vectors and 778K+ keyword documents with RRF, HyDE, iterative multi-hop retrieval, sibling expansion, recency reranking, and ReAct orchestration across databases, web search, and sandboxed Python.

1M+ vectors778K+ docs8 custom tools

Production-grade optimization

Performance & Business Impact

Improved enterprise benchmark quality and made the system fast enough for real-time assistance by cutting latency from 137 seconds to 6.6 seconds and increasing import throughput by 260x.

98.9% accuracy95% latency reduction260x import throughput

Code and recruiting automation

Autonomous Process Agents

Built agents for high-volume operational workflows, including an AI code-fix agent that resolved 12,000+ static analysis issues and an AI recruiter that derives screening criteria from job descriptions and generates one-click shortlists.

12,000+ issues resolvedJD-to-criteria extractionOne-click shortlist

Model selection and serving

LLM Infrastructure & Evaluation

Benchmarked 40+ foundation, embedding, and reranker models, then deployed optimized inference on H100/H200/B300-class GPU clusters using vLLM, SGLang, and llama.cpp.

40+ modelsGPU cluster inferenceEmbedding + rerank eval

Human + agent collaboration infrastructure

AI Collaboration Documentation Standard

Designed and open-sourced a governance standard for AI collaboration docs, separating CLAUDE.md, AGENTS.md, .cursorrules, lesson_learned, and ADRs to prevent rule sprawl, duplication, and context pollution.

Layered governanceSubmodule reuseOne-command init

Personal production product

WeChat AI Agent Platform

Launched an AI assistant inside a closed messaging ecosystem with multimodal chat, intent routing, web search, finance data, deep research, group analytics, message relay, moderation bots, and a REST API gateway.

Live usersGroup analyticsComposable API gateway

Ambient intelligence prototype

Edge AI Voice Assistant

Deployed a 24/7 Raspberry Pi 5 assistant with full-duplex realtime voice, persistent memory, calendar and reminder tools, quiet-hours scheduling, and camera-based physical context awareness.

Runs 24/7Realtime voiceLocal sensing

Superagent Deep Dive

The product is a private AI answer engine, not another chatbot.

Superagent turns fragmented enterprise knowledge into reliable AI workflows, with real deployment experience inside an S&P 500 enterprise. It is designed for domains where answers require multiple evidence types: documents, web pages, product catalogs, support tickets, structured databases, and live operational tools.

The key difference is control. Each layer is replaceable: crawler, parser, chunker, embedding model, vector database, keyword engine, reranker, LLM, tools, and UI. That makes the platform useful for teams that need private deployment and measurable answer quality instead of opaque SaaS behavior.

Request path

Source capture

Document conversion

Semantic chunking

Embedding + indexing

Hybrid search

Rerank + expand

Tool/data injection

Answer with trace

Knowledge ingestion

Crawl websites, parse PDFs and Office files, normalize HTML tables, enrich chunks with context, and keep indexes fresh as source data changes.

Hybrid retrieval

Combine semantic vector search, keyword search, RRF fusion, reranking, metadata filters, and parent/sibling expansion so answers come from the right evidence.

Agent runtime

Expose RAG and tool-using agents through an OpenAI-compatible API, with task-specific models, structured tool calls, and a single chat/completions entry point.

Structured context

Inject live product, inventory, ticket, CRM, SQL, or operational data into the answer path instead of relying only on static documents.

Evaluation loop

Trace every retrieval and generation step, compare answer quality, inspect failed recalls, and improve chunking, ranking, prompts, and tools over time.

Private deployment

Run inside a customer-controlled environment with replaceable models, vector stores, search engines, and data connectors.

OrgMem Deep Dive

In the AI era, organizational memory cannot stay a human-only wiki.

OrgMem puts organizational knowledge, decision history, and agent execution memory into one auditable system. It is both a knowledge base and a long-term memory layer: humans keep governance, while agents gain structured read/write access and continuous learning.

Memory evolution path

Episode → Pattern → Strategy → Playbook

Failures are remembered, successes are reinforced, and repeatedly verified experience becomes durable organizational know-how.

Markdown + Git as Source of Truth

Knowledge is not locked in a database or SaaS dashboard. Humans, agents, editors, and Git history can all read, edit, diff, blame, branch, and review it.

AI Writes, Humans Review

Agents can create and update knowledge, but formal organizational memory goes through staging and human review before commit, reducing hallucinated memory and poisoning risk.

Hybrid Search Beyond Vectors

Qdrant vectors, Elasticsearch BM25, RRF fusion, reranking, and entity boost combine semantic recall with exact keyword precision for real organizational knowledge.

Agent-Native Runtime Memory

OrgMem stores episodes, failure patterns, strategies, preferences, summaries, and promoted know-how so agents remember execution traces, root causes, and effective tactics.

Hot / Warm / Cold Memory

Session-local hot memory is zero-latency, project strategies preload at session start, and cross-project cold knowledge is retrieved only when needed to reduce context pollution.

Lifecycle + Promotion Pipeline

Confidence, access counts, supersession, and decay fight memory rot. Repeatedly verified patterns can be promoted into durable playbooks and lessons.

Capability Map

Built from real systems, not slideware.

Private knowledge ingestion from web pages, PDFs, Office files, email, databases, and APIs

Hybrid retrieval with vector search, BM25, RRF fusion, reranking, and structured data injection

Agent orchestration with tool calls, long-running workflows, memory, and observability

Self-hosted deployment patterns for teams that cannot send proprietary data to generic SaaS

Operational playbooks for evaluation, tracing, rollback, prompt/version control, and data refresh

Research and production track across deep research, code agents, memory systems, AI collaboration governance, recruiting automation, WeChat assistants, edge voice, and GPU inference

Commercial Offer

Start with one painful workflow. Grow into the AI layer for the company.

Pilot

2-4 weeks

Turn one messy knowledge domain into a private AI assistant with measurable answer quality.

Team Platform

6-10 weeks

Add multiple data sources, role-based workflows, review queues, and production observability.

Embedded Runtime

Custom

Use the Superagent runtime behind your own product, portal, support flow, or internal ops stack.

Have a private knowledge workflow worth automating?

Bring one workflow, one dataset, and one success metric. We will turn it into a working pilot before expanding the platform.

Start the conversation