ASTRA-X Project Demonstration
A local-first AI operations planning service that keeps scoped goals, retained feedback, generated plans, runtime evidence, and evaluation results in one traceable loop. The public surface demonstrates how a plan improves after feedback, while the repository backs that demonstration with FastAPI routes, a model-router wrapper, prompt governance, guardrail checks, a static evaluation harness, scheduler controls, and PostgreSQL or SQLite-backed run history.
The project is not a one-shot chatbot wrapper. It models planning as an operational loop: define goals by scope, capture human feedback, generate a plan, preserve the run, parse reviewable sections, and use the retained context to produce a sharper next iteration.
Public routes: /, /demo, /api/demo/frozen-runs, /api/operations/evidence, /api/evals/run, /api/health, and /ui/overthinker.html.
Planner execution passes through a routing wrapper that records provider metadata, configured and effective models, latency, token estimates, request status, and a prompt-version field for future wiring. Prompt templates are versioned with required-variable contracts before rendering, and guardrail checks record sensitive input patterns, prompt-injection phrases, empty outputs, and missing plan sections.
Core files: overthinker/services/model_router.py, overthinker/services/prompt_registry.py, overthinker/services/guardrails.py, and overthinker/services/evals.py.
| Artifact | Purpose |
|---|---|
data/private/operations/prompt_registry.json | Versioned prompt definitions, purpose metadata, active status, template body, and required variables. |
data/private/operations/llm_call_log.jsonl | Model-router call evidence with request metadata, provider state, latency, token estimates, and a reserved prompt-version field. |
data/private/operations/guardrail_events.jsonl | Input and output guardrail results for reviewable planning runs. |
data/private/operations/eval_results.jsonl | Static planning-suite pass/fail records produced by the local evaluation endpoint. |
evals/suites/planning_basic.jsonl | Repository-owned evaluation cases used to verify planning output structure and expected terms. |
The storage layer supports SQLite fallback and PostgreSQL operation. The primary schema tracks goal documents, goal items, feedback, planning runs, and parsed run sections so review state survives beyond a browser session.
The portfolio page stays credential-free. It shows frozen iteration playback and documentation, while the operator console remains separate for goals, feedback, scheduling, configuration, and manual run control.
This phase deliberately avoids overstating cloud LLMOps infrastructure. The implemented value is the inspectable AI-operations core: router metadata, prompt contracts, guardrails, static evals, feedback memory, and run evidence.