AI-native recruiting operating system targeting GSI/FSI cleared contract staffing and enterprise sales/SE recruiting.
Status: under active construction. Built from
~/.openclaw/agents/sapor/memory/PLATFORM-BLUEPRINT.md(the canonical product spec). Seedocs/SPRINT-STATUS.mdfor what is done, what is stubbed, and what requires real third-party credentials to enable.
services/ Python FastAPI services + Temporal workers (one dir per agent / domain)
rules/ Sprint 2 — OPA-backed rules engine API
screening/ Sprint 3 — screening agent
outreach/ Sprint 4 — outreach + close protection
pipeline/ Sprint 5 — Temporal pipeline orchestration
client-advisory/ Sprint 6 — client advisory + client development
interview/ Sprint 7 — voice/chat interview agent
capture/ Sprint 8 — pre-award capture intelligence
outcomes/ Sprint 9 — outcome loops + predictive ML
bench/ Sprint 10 — bench management + compliance
market/ Sprint 11 — market intelligence + data products
packages/ Shared libraries
schemas/ Pydantic + TypeScript shared schema definitions
rules-sdk/ Client SDK for the rules service
audit/ Audit-log writer (ClickHouse)
events/ Kafka/Redpanda event publisher + consumer helpers
llm/ Model router + prompt-cached Anthropic client
apps/
command-center/ Sprint 6 — Next.js client/internal dashboard
candidate-portal/ Sprint 12 — Next.js candidate-facing portal
infrastructure/
docker/ docker-compose.yml for local dev (PG+pgvector, ClickHouse, Redis, Temporal)
terraform/ AWS production stack (RDS, ElastiCache, ECS Fargate, ALB, Secrets, alarms)
k8s/ Kubernetes manifests (alternative deployment)
migrations/ SQL migrations for ClickHouse
schema/ Prisma schema (PostgreSQL operational DB)
rules/ OPA policy bundle (.rego files + tests)
docs/ Architecture decisions, sprint status, runbooks
- Node 20+, pnpm 9+
- Python 3.12+, uv
- Docker + Docker Compose
- (For OPA dev)
opaCLI:brew install opa
make bootstrap # installs all deps, brings infra up, runs migrations, generates clients
make test # runs all test suites
make dev # starts every service in dev mode (uses overmind/foreman)Each services/<name> and apps/<name> has its own pyproject.toml / package.json and can be run standalone. Run make help for the catalogue.
| Layer | Tech |
|---|---|
| Operational DB | PostgreSQL 15 + pgvector |
| Analytical DB | ClickHouse |
| Workflow | Temporal |
| Streaming | Redpanda (Kafka-compatible) |
| Rules | Open Policy Agent (OPA) |
| Backend | Python 3.12 / FastAPI |
| Frontend | Next.js 14 + React 18 + Tailwind |
| ML | scikit-learn (Sprint 9), Feast (feature store) |
| LLM | Anthropic Claude (model router: Opus / Sonnet / Haiku) with prompt caching |
| Auth | Clerk-shaped adapter (works with Clerk; mock for dev) |
| SMTP (free) primary, SendGrid adapter for production | |
| SMS | Twilio adapter (no free tier — interface only by default) |
| Voice AI | Vapi/Retell adapter interface (no free tier — see docs/INTEGRATIONS.md) |
| Object storage | S3-compatible (MinIO in dev) |
- Every AI decision is audit-logged to ClickHouse with input summary, decision, reasoning, model, confidence — see
packages/audit. - Rules engine handles process; LLMs handle judgment. Every endpoint that calls an LLM first runs deterministic guards. See
services/rules. - Candidate ownership (exclusivity, RTR, DNC, non-compete) is enforced before any submission action —
services/rulesownership policies. - Human escalation triggers are first-class — VP+ candidates, $350K+ comp, mixed signals, etc. — never silenced.
- EEOC/OFCCP traceability: every screen / submit / offer recommendation has a permanent reasoning record.
See docs/SPRINT-STATUS.md for per-sprint completion + what requires real credentials to graduate from interface-only to production.