Tileward is an AI infrastructure layer that reduces inference costs and enforces governance for large language model applications. The platform compresses models to run on existing GPUs, trims conversation context before it reaches the model, and applies policy-based guardrails before generation, all through a single API endpoint compatible with 28 clients.
Funding
Funding not disclosed
Founders
Product
Problem
Large language model applications face rising inference costs from sending full conversation histories with every request, and organizations struggle to enforce content policies consistently across their AI deployments. Running capable models requires expensive GPU infrastructure, while governance controls are often bolted on after the fact rather than built into the request path.
Solution
Tileward provides a single API layer that sits between applications and language models, reducing what gets sent to inference and controlling what the model is allowed to do. The platform compresses models—including a mixture-of-experts build that activates roughly 3B of 35B parameters per token—so they run on existing GPUs, and it trims conversation context so only relevant tiles are sent with each request. Governance policies are bound to API keys and enforced before generation, with a guard that classifies messages across 64 regulated topics plus a catch-all, recording allow-or-refuse decisions in an audit trail that stores no message text. The platform offers both hosted and self-hosted deployment, with the self-hosted option keeping all content in the customer's own VPC.
Target Audience
Primary customers are engineering teams and organizations running LLM-based applications that need to control inference costs, enforce content governance, or deploy models on their own infrastructure. The platform serves AI product companies, enterprises with data privacy requirements, and teams standardizing on a single API layer across multiple AI clients.
Features
- Model compression technology that reduces active parameters per token from 35B to ~3B in the mixture-of-experts build, enabling larger models to run on existing GPUs
- Context management that sends only the conversation tiles relevant to the current request, reducing tokens sent to the model
- Governance guard that classifies messages across 64 regulated topics plus a catch-all (65-way classification) with 94.8% topic recognition accuracy
- Per-tile locking for fine-grained control over which knowledge domains the model can access
- Audit records storing timestamps, account identifiers, actions, and verdicts without prompt or completion text, retained for a configurable window
- Single API key works across 28 clients including chat platforms, editors, agent frameworks, and gateways
- Self-hosted deployment option where the model, context store, and audit records live entirely in the customer's cloud account
- Policy enforcement bound to API keys rather than prompts, with per-device revocation and project-scoped key separation