Platform · Architecture
How logcat.ai works.
The substrate beneath every investigation. What models we use, how they're grounded, where customer data goes, and which architectural principles are non-negotiable.
Frontier models, specialized for OS-layer reasoning.
The engine routes per-task to whichever frontier model is best at the work. Long-context for whole-bugreport investigation; faster smaller models for parsing and classification.
Long-context investigation
Million-token context windows make whole-bugreport reasoning tractable for the first time. Cross-subsystem causal chains hold together across the whole artifact instead of being chunked away.
Format-agnostic profiling
For any plain-text log (vendor diagnostic streams, custom drivers, in-house tooling) the profiling layer generates domain context dynamically. For structured binaries (Android bugreports, CAN bus traces, modem QXDM) there's a native parser.
Task-routed model selection
The engine routes per-task: heavyweight long-context for investigation; smaller faster models for parsing, classification, and synthesis. No single model carries the whole load.
Grounded, verified, human-approved.
Architectural principles, not feature flags. Every claim ships with a citation; every hypothesis is verified; no AI-proposed change lands without a human in the loop.
Grounding and citation discipline.
Every claim the engine makes cites the source: the exact log line, dmesg session, bugreport section, or device tree node it came from. Findings without citations don't ship. Failed hypotheses surface as failed, not hidden. So reviewers know what was ruled out and why.
Hypothesis verification.
The engine forms hypotheses, gathers evidence, and self-corrects when the evidence pushes back. A hypothesis that fails verification is reported as failed. Not hidden, not silently dropped. The audit trail shows what was tried and why each candidate was kept or rejected.
Mandatory human-approval gates.
An architectural principle, not a feature flag. Diagnose surfaces cited findings; the engineer decides what to do with them. In research preview, Remediate's AI-proposed patches land under the same rule: the model never commits, the human always holds the merge button.
Where your data lives. What we don't do with it.
Encrypted at rest and in transit. Retained briefly by default. Never used to train AI models — ours or any provider's.
Encrypted at rest
All log data is encrypted using AES-256. Per-tenant storage isolation. No cross-tenant access.
Encrypted in transit
All data transfer uses TLS 1.3. API endpoints enforce HTTPS-only connections.
Automatic deletion
Log data is deleted after 90 days by default. Enterprise plans support custom retention policies.
No training on customer data. Ever.
Customer logs are processed in memory for the duration of an investigation and discarded when it completes. They are never used to train, fine-tune, or improve any AI model, neither logcat.ai's nor any provider's. The zero-training rule is contractual with our frontier-model providers via zero-data-retention agreements, and architectural in our pipeline.
In memory only
Log content is processed in memory by frontier-model providers and never persisted on their side.
Contractually enforced
Zero-data-retention agreements with every frontier-model provider the engine routes to.
No third-party sharing
Customer data never reaches any service beyond the model API. No vendors, no analytics, no telemetry pipelines.
Deployment shape
Self-hosted deployment architecture.
For regulated customers, logcat.ai runs entirely inside your own environment: BYOC (bring-your-own-cloud) in your VPC, or on-prem in your datacenter. Logs, indexes, embeddings, and inference all stay put; logcat.ai retains only the license and support channel.
Application tier
Console · API · Analysis workers
Data tier
Logs · Indexes · Embeddings · Audit
Customer-provided services
Inference · Object storage · Identity · SMTP
Logs stay put
Raw uploads, indexes, and vectors live only in the customer environment.
Inference stays put
LLM calls go to the customer's own frontier-model endpoint inside their VPC.
Credentials stay put
logcat.ai never holds keys to customer LLM, storage, or identity services.
- Backup orchestration
- Audit logging exportable to SIEM
- Role-based access control
- OIDC / LDAP single sign-on
- Externalized secrets management
- Signed license validation
Infrastructure
AWS-hosted
Infrastructure hosted on AWS with enterprise-grade availability and redundancy.
Tenant isolation
Isolated per-tenant storage ensures no cross-tenant data access. Each organization's data is logically separated.
Dedicated instances
Enterprise plans can deploy on dedicated cloud instances for additional isolation and performance guarantees.
Compliance
SOC2 Type II
SOC2 Type II certification is in progress with a target completion of Q4 2026.
SSO / OIDC
Enterprise plans include SSO/OIDC authentication for centralized identity management.
Custom retention
Configure data retention policies to meet your organization's compliance requirements.
Bring this in front of your security and legal team.
One page. Models, grounding, data residency, self-hosted (BYOC / on-prem) deployment, and compliance posture. The procurement-ready story for security and legal review.