AI/ML●Mid
Optimal model routing directly in Claude, Codex and Cursor
LLM routing layer when LiteLLM and Portkey already handle multi-model failover and cost optimization.
adchurch
4024d ago
Back to browse
GitHub Repository
Measuring which LLM tier (foundation / instruction-tuned / SLM) you actually need for Florida Building Code lookups — with and without RAG grounding. Reproducible benchmark harness + citation-hallucination analysis. Built as a hands-on tutorial series on modern AI system design.
0 starsJupyter Notebook
A model-routing benchmark – the routers optimize the wrong axis
by yadellopez·Jul 18, 2026·1 point·0 comments
AI Analysis
●●●BangerBig BrainDark Horse
Proves cheap local models with RAG beat cold flagship models on citation accuracy.
Strengths
- •Deterministic citation-match metric exposes hallucination rates better than vague LLM judges.
- •Data shows grounding lifts a 3.8B local model to outperform cold Opus by double.
- •Challenges the industry default of model-routing before retrieval augmentation.
Weaknesses
- •Domain specificity to building codes limits immediate generalization to other verticals.
- •Benchmark harness requires manual setup of local Ollama instances for full reproduction.
Category
Target Audience
AI engineers building RAG pipelines and legal-tech applications
Similar To
RAGAS · RouteLLM · NotDiamond
Similar Projects
AI/ML●Mid
Phase Router – capacity-aware routing for MoE
Specialized routing logic for MoE models without a demo or benchmarks.
Niche Gem
TSltd
512mo ago
AI/ML●●Solid
Smart model routing directly in Claude, Codex and Cursor
Drop-in endpoint that cuts AI coding costs 40-70% with sub-50ms routing.
Solve My ProblemSlick
adchurch
21611321d ago
AI/ML●●●Banger
role-model, a router for hybrid local/cloud AI
Protocol-first routing contract beats ad-hoc LiteLLM configs for hybrid AI deployments.
Big BrainZero to One
try-working
3219d ago
Developer Tools●●Solid
API router that picks the cheapest model that fits each query
Komilion turns model sprawl into a cost-control layer you drop in by swapping a base_url: requests are classified (regex fast path + tiny LLM) and matched to ~400 models so cheap models handle the easy stuff and premium models only run when needed. The ~60% zero‑call regex fast path and benchmark-driven routing (LMArena) are clever, pragmatic moves; the hard questions left are model-quality drift across providers and how routing decisions map to real-world user satisfaction.
Solve My ProblemWizardrySlick
robinbanner
115mo ago
AI/ML●●Solid
Apodex-1.0 – Deep research with independent verifier (90.3 BrowseComp)
90.3 BrowseComp score with verification-centric model architecture.
Niche Gem
wuqiaocauc
101mo ago