G-04 // ORBITBENCH
Multi-agent prompt & workflow evaluation workbench for side-by-side LLM latency, schema adherence, and cost benchmarking.
Production AI systems fail silently when prompt edits break downstream JSON schemas or triple token latency. OrbitBench runs parallel regression suites across Gemini, Claude, OpenAI, and open-weight models, scoring strict JSON Schema compliance, tool-call determinism, time-to-first-token, and cost-per-1k runs.
Born directly inside OXAKALABS while debugging multi-agent hackathon harnesses under 48-hour pressure, OrbitBench replaces vibes-based prompt tweaking with repeatable engineering telemetry.
Parallel Model Harness: Concurrent execution runner with unified token, latency, and cost accounting across 6 model providers.
Strict Zod/JSON-Schema Diffing: Highlights exact field-level structural drift and hallucinated keys across 100+ test fixtures.
CI/CD GitHub Action Gate: Blocks pull requests when schema compliance drops below 99.5% or p95 latency spikes.
Benchmarked Gemini 2.5 Flash structured outputs against multi-provider hackathon extraction fixtures.
TECHNICAL STACK & RUNTIME
High engineering mass & commercial pull confirmed. Assign a G-XX product designation, allocate dedicated Labs sprint cycle, and deploy waitlist micro-site.
Receive release binaries, sandbox credentials, and architectural updates for G-04.
AI interactive pitch & launch studio converting founder notes and architecture docs into kinetic web micro-sites.
Autonomous WhatsApp, email, and dashboard operations agent for high-volume SMEs (intake, quoting, invoicing, daily briefs).
Browser-based AI brand & motion asset lab turning product URLs and screenshots into minimalist-luxury launch teasers.