Everything your technical team needs to evaluate HireBuddha. No sales pitch — just engineering.
SSL termination and request routing at the edge.
The generative, adaptive UI assembled per user and task.
The orchestration core. Async throughout.
Capability requests and WebSocket streaming to the engine.
Realtime audio for OmniBuddha, Gemini Live, and ChatGPT Realtime.
Relational state plus 768-dim semantic search.
Rate limiting and the async job queue for child dispatch.
A single hierarchical_entities table holds all five tiers — Loop, Process, Agent, Skill, Action — each with identity, logic gate, planning, capabilities, governance, I/O contract, and per-entity router prefs.
An 8-stage control flow — Perceive, Strategize, Pre-Critic, Act, Observe, Post-Critic, Reflect, Decide — with a circuit breaker after 3 consecutive pre-critic blocks and cost tracking per step.
Parent state is serialized to a snapshot, children run on the queue, and a resume worker rehydrates the parent. Cost aggregates across the whole hierarchy.
Working, episodic, semantic (pgvector, cosine > 0.70), and CORTEX tree memory with auto-checkpointing at 8,000 tokens and viewport slicing.
No agent calls a model SDK directly. It requests a capability and the Intelligent Router binds it — BabyBuddha plus Claude Opus, ChatGPT, Gemini, GLM, Qwen, and Mistral.
Per-tenant allow-lists and data-residency filters decide which providers are eligible before scoring. Pin a model where required; full attribution and audit of every model that touched data.
Tool synthesis and self-healing run sandboxed, test-gated, versioned, and rollback-safe, with optional HITL on promotion. Production is never edited in place.
Harvests reflections, KPIs, critic verdicts, router performance, and HITL feedback, then runs a per-tenant optimizer that tunes charters, routing, tools, and the schema.