1. Zero-Retention Default Gateway
Flexlora operates with a strict default zero-retention edge architecture. Your prompt payloads, completions, and embeddings are streamed in-memory through our edge Workers and are never persisted to disk or database during standard operations, unless Tier-2 debug payload logging is explicitly configured.
2. Transmission to Upstream Inference Providers
To execute machine learning inference and generate completions, Flexlora securely transmits your prompts, context messages, and generation parameters over encrypted TLS connections to authorized third-party upstream inference providers and model infrastructure subprocessors. Transmission is confined strictly to fulfilling real-time inference requests.
3. Multi-Tier Data Storage & Payload Retention Policies
Flexlora maintains a transparent, strictly segregated multi-tier data storage architecture:
- Tier 1 — Usage Metrics & Accounting (Supabase PostgreSQL): Aggregated request telemetry and billing records are stored in our encrypted relational database (Supabase PostgreSQL). Recorded data includes request timestamps, requested public model identifier, token usage (input, output, and total), latency metrics, HTTP status codes, API key identifiers, and credit deductions. Tier 1 records contain quantitative operational metadata only; prompt inputs and completion texts are never written to Tier 1 storage.
- Tier 2 — Payload & Debug Archiving (Cloudflare R2): Full request prompts, model completions, and raw error response bodies are archived only when debug payload logging is activated or when sampled for security auditing and edge anomaly diagnostics. These encrypted payload archives are stored in Cloudflare R2 object storage (
flexlora-logs) under a strict 30-day automated rolling lifecycle policy, after which they are permanently deleted.
4. Information We Collect
We collect basic account credentials (email address, full name, company name) and connection metadata (client IP address for sliding-window rate limiting, DDoS defense, and abuse prevention).
5. No Model Training on Customer Data
Under our enterprise provider agreements and governance terms, neither Flexlora nor our upstream foundation model providers utilize your private prompts, completions, embeddings, or fine-tuning datasets to train, retrain, or improve artificial intelligence foundation models.
6. Payment Information
For manual bKash deposits, we store the sender phone number and Transaction ID (TrxID) solely for financial accounting and deposit verification.