How experientiallabs/experiential works
Experiential combines an OpenAI-compatible model gateway with tooling that turns agent traces into evaluated and optimized routers. The main workflow is: capture or ingest traces, build project data, simulate and evaluate candidate behavior, then fit and serve a router. Separate branches cover gateway setup and governance, judge calibration, and model fine-tuning. The graph emphasizes documented workflows and relationships visible in the supplied source index; many implementation files were not sampled, so uncertain internal calls are omitted.
Main entry points: routing.py Request routing · server.rs Native API server · store.py Project artifacts · workflow.py Router fitting
Components 22
Gateway and runtime
- CLI
exp/cli/app.py - Gateway setup
exp/cli/gateway/app.py - Gateway application
exp/runtime/router/application.py - Native API server
exp/runtime/gateway/native/src/server.rs - Request routing
exp/runtime/gateway/routing.py - Model catalog
exp/common/models/catalog.py - Model providers
exp/runtime/models/providers/transport.py
Traffic and projects
- Traffic capture
exp/runtime/capture/policy.py - Trace ingestion
exp/common/traces/ingest - Project artifacts
exp/common/project/store.py - Trace mining
exp/simulation/mining/service.py - World model
exp/simulation/world_model/application.py
Evaluation and optimization
- Evaluation service
exp/optimize/evaluation/service.py - Router fitting
exp/optimize/router/fit/workflow.py - Automatic optimization
exp/optimize/router/automatic/service.py - Router runtime
exp/runtime/router/runtime.py
Judging and training
- Judge commands
exp/cli/judge/app.py - Judge calibration
exp/optimize/router/judging/service.py - Review persistence
exp/optimize/router/judging/review.py - Model fine-tuning
exp/optimize/model/sft/training.py
Other components
- Developer
- Model provider
Connections 23
- Developer → CLI: runs commands
README.md - CLI → Gateway setup: configures gateway
exp/cli/app.py - Developer → Native API server: sends API request
README.md - Gateway application → Native API server: starts data plane
exp/runtime/router/application.py - Native API server → Request routing: routes requests
exp/runtime/gateway/native/src/server.rs - Request routing → Model catalog: resolves models
exp/runtime/gateway/routing.py - Request routing → Model providers: dispatches requests
exp/runtime/gateway/routing.py - Model providers → Model provider: calls inference
- Developer → Traffic capture: captures traffic
README.md - Developer → Trace ingestion: uploads traces
README.md - Trace ingestion → Project artifacts: stores trace data
- Developer → Trace mining: builds project
README.md - Trace mining → Project artifacts: persists project artifacts
- Project artifacts → World model: supplies project data
- Developer → Evaluation service: evaluates behavior
README.md - Evaluation service → Router fitting: provides evaluation
- Developer → Automatic optimization: optimizes router
README.md - Automatic optimization → Router fitting: fits router
exp/optimize/router/automatic/service.py - Router fitting → Router runtime: produces router
- Developer → Judge commands: calibrates judge
README.md - Judge commands → Judge calibration: runs calibration
exp/cli/judge/app.py - Judge calibration → Review persistence: records reviews
exp/optimize/router/judging/service.py - Developer → Model fine-tuning: fine-tunes model
README.md