Instrumentation SDK
Emits structured trajectory events — messages, tool calls, outcomes, tokens — from your existing agent harness.
≤5 lines integration • zero behavioral impactReflex separates execution into two strict operational loops joined by a single choke point — the Adapter & Model Registry. Online serving never triggers training; offline training never serves live traffic without a promotion pass.
Scroll to traverse the full pipeline — from instrumentation to serving.
Emits structured trajectory events — messages, tool calls, outcomes, tokens — from your existing agent harness.
≤5 lines integration • zero behavioral impactVersioned, customer-owned storage. Configured PII categories masked before any dataset eligibility.
≥1M traces/month • build blocked if redaction failsSuccess filtering, near-duplicate clustering, and stratified sampling across rare task slices.
Auditable manifest per buildOne-command training producing versioned QLoRA adapters or standalone vertical SLMs.
Single 48 GB GPU • full lineage recordedCandidates evaluated on frozen production-derived traces with deterministic checks and pass-k reliability.
Emits PASS / REVIEW / FAILVersion lineage and lifecycle state. The sole path to production — activation refused without a passing comparative report.
THE CHOKE POINTOpenAI-compatible endpoint serving a shared base plus adapters or standalone SLMs with continuous batching.
≤5% p95 latency overheadConfidence-threshold routing strips baked stable prompt segments and monitors live metrics.
Automatic reversion in <60 secondsA hard six-step promotion protocol against a frozen production Golden Set. Only a PASS verdict permits activation.
Candidate adapter or SLM registered with exact dataset hash, base model ID, and hyperparameters.
The candidate runs against the frozen golden set. Infrastructure failure defaults to “inconclusive” with zero state change.
Per-slice report compares the candidate against both the prompted baseline and the currently active model version.
Only a PASS verdict permits activation. REVIEW or FAIL verdicts are blocked absent an explicit, audited human override.
Passing artifacts hot-load into the serving runtime while arming the rollback monitor with the report’s regression thresholds.
If live metrics breach thresholds, traffic reverts to the prior version in under 60 seconds — or to the terminal frontier-model fallback.
The Prompt Slimming Engine removes stable segments at runtime. Compute your break-even volume directly from your existing API invoices — before a single training job.
Adapters are disposable artifacts regenerated via one-command retrain loops. Your durable, compounding asset is your proprietary trajectory corpus — and it stays yours.
The customer-owned data plane is a hard design goal: trajectories, datasets, and weights never leave customer infrastructure.
Training and serving deploy into your cloud project. Trajectory stores, PII-redacted datasets, and model registries stay in your storage buckets.
Fully disconnected operation for regulated environments. No external API calls, no telemetry, no egress — ever.
Reserve single 48 GB-class accelerators for predictable throughput without managing hardware or sharing capacity.
STACK // Open-weights Gemma-class foundations fine-tuned as standalone vertical SLMs or served as frozen bases in the multi-LoRA runtime • frontier API operates as the initial prompted baseline and terminal fallback • single 48 GB-class GPU instances handle both the training pipeline and continuous-batching serving, in-region.
A brief shadow deployment of the SDK answers break-even before any training job.