Two-tier orchestration: proprietary APIs plan and verify; local Ollama models handle small coding task contracts. Data-driven capex decision after 1–2 months of measured MVP.
- Orchestrator tier: API (Claude Opus/Sonnet, or alternatives) — decomposes, plans, reviews, merges
- Worker tier: Local Ollama models — executes small, constrained coding tasks with minimal repo context
- Goal: Replace 70%+ of API grunt spend with local inference without losing >15% quality
- Decision gate: 1–2 months of real data → stay / upgrade / hybrid
git clone https://github.com/AdarGit008/local-ai.git
cd local-ai| Path | Purpose |
|---|---|
configs/ |
Worker profile definitions (model, context, task types) |
data/ |
Benchmark results, hardware tests, session summaries |
docs/ |
Project intent, prior orchestration research |
mvp/ |
MVP architecture, task contracts, instrumentation harness |
research/ |
External landscape synthesis, hardware capability assessment |
- Parallel workers are viable: Ollama runs different models as separate GPU processes; same-model concurrency possible via
OLLAMA_NUM_PARALLEL - Best 6GB model candidate: GLM-4.7-Flash (30B MoE, 3B active, ~5GB Q4, 59.2% SWE-bench)
- Industry benchmark: EvalPlus (HumanEval+, MBPP+) gives comparable pass@1 scores
Derived — see docs/start-here.md.