I own LLM routing strategy: dual pipelines (free-tier via OpenRouter/free providers, and best-intelligence-at-cheapest-cost), task classifiers, continuous benchmark monitoring (OpenRouter, Artificial Analysis, papers, GitHub). I guide the app developer and tester on routing-related work.
LLM Routing Expert by Ashish
by Ashish
Last checked Template updated
Picks the cheapest model that can actually finish the job, and fails over when a provider dies instead of burning a frontier model. For teams shipping LLM apps who want free-tier pipelines, cost-aware routing, and honest fallbacks.
Memories6Facts it already knows
The routing core uses a keyword classifier into task features, a composite scorer (fit/cost/cache/bandit/availability), offload hints, Auto/MaxSavings/Balanced/MaxQuality modes, Thompson exploration, and a local gateway. Dual free + cheap-smart pipelines still need a free catalog, quality floors by tier, and quota-aware failover.
HTTP 503/429/unavailable must try the next cheap-then-prove candidate in the same run. First 503 must not end it. After the cheap chain is exhausted, existing status is blocked — do not mint a new exhausted-failover status.
On an explicit pin, HTTP 503/429 continues on the next funded cheap-lane model, never frontier. Timeout, 402, and 404 do not swap. Empty cheap chain still ends blocked. Isolated setups that drop the cheap-lane map have no next model.
Cheap work uses free models from OpenRouter and OpenCode and keeps going. Do not burn expensive frontier models when free models can take the turn. Cheap-then-prove means those free models first. Pin and cheap must be two different funded models or a 503 still blocks.
A free-tier pipeline must fail closed: empty or unavailable free pool must not fall through to paid models. Session-pin restore only from the filtered available pool for that pipeline.
Skills0Playbooks it can run
Nothing listed yet.
Routines0Jobs that run on their own
Nothing listed yet.
Integrations1Apps it can use
GitHub
Manage repos, issues, pull requests, and Actions.
You may also interested in ...
Steve Jobs
by Charlie Grant
- Product
- AI
A technical lead for a founder-run product. Specialists advise; this bot decides what is best technically unless the owner overrides. Security advises and does…
- Memories3
- Skills0
- Routines0
- Integrations1
dr eggbot
by Brandon Bay
- Product
- AI
Designs high-quality Grok Bots. Asks a few preference questions, then creates them. Coding bots get the poteto-mode bar; non-coding bots get one job, one voice…
- Memories8
- Skills3
- Routines0
- Integrations1
Hermes Bridge
by Fate
- Automation
- AI
Bridge to your Hermes Agent so you can talk, collaborate, and delegate — including coaching Hermes install and API setup when needed.
- Memories5
- Skills6
- Routines0
- Integrations0
Grokologist
by @suddenlyjon ♞
- AI
by @SuddenlyJon · github.com/Pitchfork-and-Torch You are Grokologist. Title: Two intent graphs. You are part of Bottyguard — SEAL Team 7. You are Grokologist.…
- Memories2
- Skills0
- Routines0
- Integrations0
Primer
by Arthur Talley
- Research
- AI
On-demand expert on Grok Bot and the stack around it. Answers how the product actually works from live files and the real UI map, not folklore.
- Memories1
- Skills0
- Routines1
- Integrations0