Gremlin
How teams standardize and automate reliability across their organization-one service at a time.
AI Sandbox
What Is the AI Sandbox?
The AI Sandbox is a platform developed by Cognipeer and operated by Hepapi in partnership with AWS. It provides enterprises with a dedicated, isolated AI environment available in the AWS Local Zone or on AWS Cloud, built around the Cognipeer Console platform.
Rather than waiting for AI infrastructure to be provisioned, configured, and secured from scratch, organizations get a ready-to-operate control plane covering the full AI stack: from model inference and vector stores to agent tracing, guardrails, and RAG pipelines.
This is not a demo environment. It is designed to run real workloads, serve multiple teams simultaneously with full tenant isolation, and scale when your needs grow, including GPU-backed inference when AWS Local Zone GPU capacity becomes available.
Why It Matters for Your Organization
AI adoption in the enterprise often stalls not because of a lack of interest, but because of infrastructure friction: Who manages the models? How do you isolate data per team? How do you ensure compliance? How do you evaluate model outputs before they reach production?
The AI Sandbox eliminates that friction. It gives you:
-
A single control plane for inference, retrieval, agent execution, and monitoring
-
Tenant-level data and credential isolation across business units or projects
-
OpenAI-compatible APIs so your development teams can start without rewriting code
-
Built-in guardrails and PII protection from day one
-
Deployment on AWS Local Zone, keeping your data within your region's borders
-
GPU-ready architecture to take advantage of future Local Zone GPU capacity
"A true HubSpot drag-and-drop web page builder! Generator is SO easy to customize without having to touch any code. Highly recommend installing this."
Platform Capabilities
The AI Sandbox is built on Cognipeer Console, a production-grade multi-tenant AI control plane. Key capabilities include:
Model Inference & Model Hub
Run LLM inference across multiple providers through a single OpenAI-compatible gateway. Register, configure, and switch models without changing your application code.
Multi-Tenant Architecture
Full data, credentials, and project isolation between teams or customers, enforced at the request level, not as an afterthought.
Vector Stores & RAG Pipelines & Memory
Ingest documents, build retrieval pipelines, and serve grounded responses. Manage prompt templates with versioning and persistent memory stores - all without assembling separate infrastructure for each component.
Agent Tracing & Evaluation
Track every agent session end-to-end. Evaluate model outputs systematically before they reach production users.
Guardrails & PII Protection
Apply input/output validation policies with keywords, regex, and LLM-based guardrails. Detect and automatically redact personally identifiable information.
GPU Fleet Management
Onboard GPU hosts, deploy models, configure pools, and load balance. Ready to scale when AWS Local Zone GPU infrastructure becomes available.
Built for AWS Local Zone and Beyond
Hepapi is an AWS Advanced Tier Services Partner. The AI Sandbox is designed to run natively on AWS, including AWS Local Zones, enabling low-latency, data-residency-compliant deployments for organizations operating across any supported region
For organizations in any Local Zone or AWS region where managed AI services such as Bedrock or AgentCore are not yet available, the AI Sandbox provides an equivalent operational surface that can be deployed locally or via AWS Cloud.
This positions Hepapi as the partner that can deliver differentiated AI infrastructure today, with a clear path to scale as AWS Local Zone capabilities expand, including GPU availability.
.png)
Who Is It For?
Technical Decision-Makers
CTOs, Technical Leads, and Platform Engineers who need a production-grade AI infrastructure without the overhead of building one from scratch.
Business & Innovation Leaders
Directors and enterprise stakeholders looking to move AI initiatives forward with a defined scope, validated environment, and a trusted delivery partner.
How It Works
Scoping & Environment Design
We align on your use cases, tenant structure, data residency requirements, and deployment target (Local Zone or AWS Cloud).
Deployment & Configuration
Hepapi deploys and configures the Cognipeer Console environment on your designated AWS infrastructure. Providers, models, vector backends, and storage are configured to meet your requirements.
Onboarding & Enablement
Your teams are onboarded to the platform. APIs are documented, access controls are configured per tenant, and guardrails are set before workloads go live.
Operate & Scale
You run your AI workloads on a stable, monitored platform. Hepapi provides ongoing support, and the environment scales as your usage and GPU availability grow.
Cognipeer Console vs. AWS AI & Agent Stack
| Area | AWS Service | Cognipeer | Assessment |
| Inference / models | Amazon Bedrock – managed inference (Claude, Nova, Llama, Mistral) | 7+ providers, including Bedrock, behind a single OpenAI-compatible API | AWS provides CP abstracts |
| RAG | Bedrock Knowledge Bases – managed ingest, embedding, reranking, source attribution | Full RAG module: chunking, vector search, reranking, metadata filters | On par |
| Vector store | OpenSearch Serverless / S3 Vectors / Aurora pgvector | Selectable backend: SQLite (default), Milvus, S3 Vectors | On par |
| Managed agent | Bedrock Agents – action groups, KB association, orchestration | Agent service with state management, context summarization, and tool-call limits | On par |
| Content safety | Bedrock Guardrails – PII, denied topics, content filters | Guardrail module: PII detection, moderation, prompt shield, custom LLM evaluator | On par |
| OCR / documents | Amazon Textract / Bedrock Data Automation | OCR API (Textract or VLM based) with bounding box and confidence output | On par CP uses Textract |
| Speech | Amazon Polly (TTS) + Transcribe (STT), separate services | Single OpenAI-compatible API for speech, transcription, and translation | On par CP unified |
| Identity (human) | Amazon Cognito / IAM Identity Center | JWT, API tokens, LDAP/SAML/OIDC support, per-tenant database isolation | On par |
| Code execution | AWS Lambda / Fargate / Firecracker microVM | isolated-vm (V8) + Docker-in-Docker with persistent volumes | AWS stronger |
| GPU / model hosting | Amazon SageMaker / Bedrock Custom Model Import | EE GPU Fleet – distributed GPU inference | On par |
| Observability | CloudWatch + X-Ray – logs, metrics, traces | Session-level tracing, health checks, alert engine | AWS more mature |
| Orchestration | AWS Step Functions – state-machine workflows | Queue-based (BullMQ/Redis) with cluster task assignment | AWS more mature |
| Batch | Bedrock Batch Inference | Batch service – bulk async job submission and tracking | On par |
| Prompt management | Bedrock Prompt Management + Flows | Versioned templates, environment-based deployment, Prompt Optimizer (EE) | Cognipeer stronger |
| Enterprise assistant | Amazon Q Business – packaged RAG assistant | Equivalent can be assembled from agent + RAG + tools; white-label ready | Different |
This document is intended for internal technical evaluation. · Cognipeer Console vs. AWS AI & Agent Stack – Section 4 (simplified)
1. Positioning at a Glance
The two products look like competitors but sit at different layers. AWS offers a horizontal infrastructure (PaaS) layer: you write the agent, AWS runs and scales it securely. Cognipeer is a vertical, integrated application platform: it bundles the entire application layer – agent framework, RAG, guardrails, red-teaming, spend tracking, and voice – into a single product.
| Dimension | Cognipeer vs. AWS |
| Positioning |
Cognipeer
End-to-end, multi-tenant AI application platform
AWS
Modular set of managed agent infrastructure services
|
| Deployment |
Cognipeer
Self-hosted / on-prem (K8s, Docker) + managed option
AWS
AWS account only, fully managed serverless
|
| Model independence |
Cognipeer
Provider-agnostic: OpenAI, Anthropic, Bedrock, Vertex, Azure, Together, Ollama
AWS
Model-agnostic but Bedrock-centric
|
| Billing |
Cognipeer
Infrastructure cost + license; predictable
AWS
Consumption-based (vCPU-hour + GB-hour); free while CPU waits on I/O
|
| Lock-in |
Cognipeer
Low – portable, open provider layer
AWS
High – tied to VPC, IAM, CloudWatch, Cedar, S3
|
In one sentence: Cognipeer's scope goes beyond AgentCore's nine services (RAG, an OpenAI-compatible API suite, prompt management, red-teaming, voice, a GPU fleet). In return, AWS goes deeper on isolation (microVM), serverless scale, and standard integrations (OTEL, Cedar policy). In many cases, the two can be complementary: AWS provides the infrastructure and security primitives, Cognipeer the application and product layer.
2. Scorecard
| Area | AWS vs. Cognipeer |
|
Runtime isolation & scale
AWS |
AWS
microVM, serverless
Cognipeer
isolate + Docker + volume
|
|
Sandbox / code execution (state)
Cognipeer |
AWS
Code Interpreter snapshot
Cognipeer
fork / persist
|
|
Memory
On par |
AWS
automatic consolidation
Cognipeer
selectable backend
|
|
Tool gateway / MCP
AWS |
AWS
semantic search, one-click connectors
Cognipeer
OpenAPI/MCP + resilience
|
|
Identity
Different |
AWS
workload vault, Cedar
Cognipeer
per-tenant DB, RBAC, LDAP
|
|
Browser automation
On par |
AWS
CAPTCHA, replay
Cognipeer
LLM extraction, MCP
|
|
Observability / tracing
AWS |
AWS
OTEL, third-party APM
Cognipeer
internal + spend + alerts
|
|
Evaluation + red-teaming
Cognipeer |
AWS
13 evaluators, live scorer
Cognipeer
OWASP red-teaming
|
|
Policy / Guardrails
Split |
AWS
Cedar declarative authorization
Cognipeer
PII / moderation / shield
|
|
RAG
On par |
AWS
Knowledge Bases
Cognipeer
full module + rerank
|
|
Ready-made client APIs
Cognipeer |
AWS
Bedrock API (proprietary)
Cognipeer
OpenAI-compatible suite
|
|
Prompt management / optimizer
Cognipeer |
AWS
Prompt Management
Cognipeer
versioned + optimizer
|
|
Speech + real-time / voice
Cognipeer |
AWS
Polly / Transcribe / Nova Sonic
Cognipeer
Enterprise real-time, unified
|
|
Inference / models
AWS provides |
AWS
Bedrock (broad catalog)
Cognipeer
provider-agnostic (incl. Bedrock)
|
|
Deployment flexibility
Cognipeer |
AWS
AWS only
Cognipeer
self-host / on-prem
|
3. When to Choose Which
- Amazon Bedrock AgentCore documentation
- AgentCore GA announcement (October 2025)
- Amazon Bedrock documentation (Knowledge Bases, Guardrails, Agents)
- AgentCore pricing page
- Cognipeer Console codebase review (Community + Enterprise editions)
This document is prepared for technical evaluation purposes.
1. Positioning at a Glance
The two products look like competitors but sit at different layers. AWS offers a horizontal infrastructure (PaaS) layer: you write the agent, AWS runs and scales it securely. Cognipeer is a vertical, integrated application platform: it bundles the entire application layer – agent framework, RAG, guardrails, red-teaming, spend tracking, and voice – into a single product.
| Dimension | AWS (AgentCore + Bedrock family) |
| Positioning |
Cognipeer
End-to-end, multi-tenant AI application platform
AWS
Modular set of managed agent infrastructure services
|
| Deployment |
Cognipeer
Self-hosted / on-prem (K8s, Docker) + managed option
AWS
AWS account only, fully managed serverless
|
| Model independence |
Cognipeer
Provider-agnostic: OpenAI, Anthropic, Bedrock, Vertex, Azure, Together, Ollama
AWS
Model-agnostic but Bedrock-centric
|
| Billing |
Cognipeer
Infrastructure cost + license; predictable
AWS
Consumption-based (vCPU-hour + GB-hour); free while CPU waits on I/O
|
| Lock-in |
Cognipeer
Low – portable, open provider layer
AWS
High – tied to VPC, IAM, CloudWatch, Cedar, S3
|
In one sentence: Cognipeer's scope goes beyond AgentCore's nine services (RAG, an OpenAI-compatible API suite, prompt management, red-teaming, voice, a GPU fleet). In return, AWS goes deeper on isolation (microVM), serverless scale, and standard integrations (OTEL, Cedar policy). In many cases, the two can be complementary: AWS provides the infrastructure and security primitives, Cognipeer the application and product layer.
2. Scorecard
| Area | AWS | Cognipeer | Standout |
| Runtime isolation & scale | microVM, serverless | isolate + Docker + volume | AWS |
| Sandbox / code execution (state) | Code Interpreter | snapshot / fork / persist | Cognipeer |
| Memory | automatic consolidation | selectable backend | On par |
| Tool gateway / MCP | semantic search, one-click connectors | OpenAPI/MCP + resilience | AWS |
| Identity | workload vault, Cedar | per-tenant DB, RBAC, LDAP | Different |
| Browser automation | CAPTCHA, replay | LLM extraction, MCP | On par |
| Observability / tracing | OTEL, third-party APM | internal + spend + alerts | AWS |
| Evaluation + red-teaming | 13 evaluators, live scorer | + OWASP red-teaming | Cognipeer |
| Policy / Guardrails | Cedar declarative authorization | PII / moderation / shield | Split |
| RAG | Knowledge Bases | full module + rerank | On par |
| Ready-made client APIs | Bedrock API (proprietary) | OpenAI-compatible suite | Cognipeer |
| Prompt management / optimizer | Prompt Management | versioned + optimizer | Cognipeer |
| Speech + real-time / voice | Polly / Transcribe / Nova Sonic | Enterprise real-time, unified | Cognipeer |
| Inference / models | Bedrock (broad catalog) | provider-agnostic (incl. Bedrock) | AWS provides |
| Deployment flexibility | AWS only | self-host / on-prem | Cognipeer |
3. When to Choose Which
- Amazon Bedrock AgentCore documentation
- AgentCore GA announcement (October 2025)
- Amazon Bedrock documentation (Knowledge Bases, Guardrails, Agents)
- AgentCore pricing page
- Cognipeer Console codebase review (Community + Enterprise editions)
1. Positioning at a Glance
The two products look like competitors but sit at different layers. AWS offers a horizontal infrastructure (PaaS) layer: you write the agent, AWS runs and scales it securely. Cognipeer is a vertical, integrated application platform: it bundles the entire application layer – agent framework, RAG, guardrails, red-teaming, spend tracking, and voice – into a single product.
| Dimension | Cognipeer Console | AWS (AgentCore + Bedrock family) |
| Positioning | End-to-end, multi-tenant AI application platform | Modular set of managed agent infrastructure services |
| Deployment | Self-hosted / on-prem (K8s, Docker) + managed option | AWS account only, fully managed serverless |
| Model independence | Provider-agnostic: OpenAI, Anthropic, Bedrock, Vertex, Azure, Together, Ollama | Model-agnostic but Bedrock-centric |
| Billing | Infrastructure cost + license; predictable | Consumption-based (vCPU-hour + GB-hour); free while CPU waits on I/O |
| Lock-in | Low – portable, open provider layer | High – tied to VPC, IAM, CloudWatch, Cedar, S3 |
In one sentence: Cognipeer's scope goes beyond AgentCore's nine services (RAG, an OpenAI-compatible API suite, prompt management, red-teaming, voice, a GPU fleet). In return, AWS goes deeper on isolation (microVM), serverless scale, and standard integrations (OTEL, Cedar policy). In many cases, the two can be complementary: AWS provides the infrastructure and security primitives, Cognipeer the application and product layer.
2. Scorecard
| Area | AWS | Cognipeer | Standout |
| Runtime isolation & scale | microVM, serverless | isolate + Docker + volume | AWS |
| Sandbox / code execution (state) | Code Interpreter | snapshot / fork / persist | Cognipeer |
| Memory | automatic consolidation | selectable backend | On par |
| Tool gateway / MCP | semantic search, one-click connectors | OpenAPI/MCP + resilience | AWS |
| Identity | workload vault, Cedar | per-tenant DB, RBAC, LDAP | Different |
| Browser automation | CAPTCHA, replay | LLM extraction, MCP | On par |
| Observability / tracing | OTEL, third-party APM | internal + spend + alerts | AWS |
| Evaluation + red-teaming | 13 evaluators, live scorer | + OWASP red-teaming | Cognipeer |
| Policy / Guardrails | Cedar declarative authorization | PII / moderation / shield | Split |
| RAG | Knowledge Bases | full module + rerank | On par |
| Ready-made client APIs | Bedrock API (proprietary) | OpenAI-compatible suite | Cognipeer |
| Prompt management / optimizer | Prompt Management | versioned + optimizer | Cognipeer |
| Speech + real-time / voice | Polly / Transcribe / Nova Sonic | Enterprise real-time, unified | Cognipeer |
| Inference / models | Bedrock (broad catalog) | provider-agnostic (incl. Bedrock) | AWS provides |
| Deployment flexibility | AWS only | self-host / on-prem | Cognipeer |
3. When to Choose Which
- Amazon Bedrock AgentCore documentation
- AgentCore GA announcement (October 2025)
- Amazon Bedrock documentation (Knowledge Bases, Guardrails, Agents)
- AgentCore pricing page
- Cognipeer Console codebase review (Community + Enterprise editions)
Tell The Reader More
Cognipeer Console vs. AWS AI & Agent Stack
Section 4 - Comparison with Other AWS Services (Beyond AgentCore). AWS services that map to Cognipeer capabilities outside
the AgentCore scope.
| Area | AWS vs. Cognipeer |
|
Inference / models
AWS providesCP abstracts |
AWS
Amazon Bedrock – managed inference (Claude, Nova, Llama, Mistral)
Cognipeer
7+ providers, including Bedrock, behind a single OpenAI-compatible API
|
|
RAG
On par |
AWS
Bedrock Knowledge Bases – managed ingest, embedding, reranking, source attribution
Cognipeer
Full RAG module: chunking, vector search, reranking, metadata filters
|
|
Vector store
On par |
AWS
OpenSearch Serverless / S3 Vectors / Aurora pgvector
Cognipeer
Selectable backend: SQLite (default), Milvus, S3 Vectors
|
|
Managed agent
On par |
AWS
Bedrock Agents – action groups, KB association, orchestration
Cognipeer
Agent service with state management, context summarization, and tool-call limits
|
|
Content safety
On par |
AWS
Bedrock Guardrails – PII, denied topics, content filters
Cognipeer
Guardrail module: PII detection, moderation, prompt shield, custom LLM evaluator
|
|
OCR / documents
On parCP uses Textract |
AWS
Amazon Textract / Bedrock Data Automation
Cognipeer
OCR API (Textract or VLM based) with bounding box and confidence output
|
|
Speech
On parCP unified |
AWS
Amazon Polly (TTS) + Transcribe (STT), separate services
Cognipeer
Single OpenAI-compatible API for speech, transcription, and translation
|
|
Identity (human)
On par |
AWS
Amazon Cognito / IAM Identity Center
Cognipeer
JWT, API tokens, LDAP/SAML/OIDC support, per-tenant database isolation
|
|
Code execution
AWS stronger |
AWS
AWS Lambda / Fargate / Firecracker microVM
Cognipeer
isolated-vm (V8) + Docker-in-Docker with persistent volumes
|
|
GPU / model hosting
On par |
AWS
Amazon SageMaker / Bedrock Custom Model Import
Cognipeer
EE GPU Fleet – distributed GPU inference
|
|
Observability
AWS more mature |
AWS
CloudWatch + X-Ray – logs, metrics, traces
Cognipeer
Session-level tracing, health checks, alert engine
|
|
Orchestration
AWS more mature |
AWS
AWS Step Functions – state-machine workflows
Cognipeer
Queue-based (BullMQ/Redis) with cluster task assignment
|
|
Batch
On par |
AWS
Bedrock Batch Inference
Cognipeer
Batch service – bulk async job submission and tracking
|
|
Prompt management
Cognipeer stronger |
AWS
Bedrock Prompt Management + Flows
Cognipeer
Versioned templates, environment-based deployment, Prompt Optimizer (EE)
|
|
Enterprise assistant
Different |
AWS
Amazon Q Business – packaged RAG assistant
Cognipeer
Equivalent can be assembled from agent + RAG + tools; white-label ready
|
This document is intended for internal technical evaluation. · Cognipeer Console vs. AWS AI & Agent Stack – Section 4 (simplified)
Ready to Run AI at Scale?
Whether you are evaluating AI infrastructure for the first time or looking to consolidate fragmented tooling into a single, governable platform, the AI Sandbox provides a production-ready starting point, deployed on AWS and operated by Hepapi.