Tegh🛡️
128 posts
@teghsm
Lead AI Engineer @Coinbase | Opinions my own | @uoft alum
Introducing Coinbase Advisor. One of the first SEC-registered AI-powered investment advisors in the world. High-quality advice - specific to you - is now one chat away. Real-time portfolio analysis, automated tax loss harvesting, and more. Rolling out to Coinbase One members.
some of the stuff we shipped in H1 on the card team @coinbase mux: an internal tool that lets you code with any harness or cli that coinbase eng have access to. mux also helps with token spend. (see @brian_armstrong’s recent post on AI spend) for example: - use one opus 4.8 session for planning and route to @cursor_ai composer 2.5 for implementation by spawning new mux sessions on different worktrees rolled it out across the company. read more here -> coinbase.com/blog/coding-ha… card agent on slack, which basically turned into another teammate: - every page or monitor alert is triaged by Card agent and it effectively builds an intelligence to help for future pages - you can ping it for investigations and it retains memory from past interactions - runs our weekly ops reviews - logs every bug, triage, page and monitor alert into linear, which doubles as its memory - if it sees a page or monitor it can fix with code, it uses our in-house software factory (Forge @chintanturakhia) to do it partnered with agentX @kylecesmat @siddharthkul in the platform org(@rwitoff) on a slack multiplexer that lets any team spin up a slack bot trivially with only one slack app. AI advisor bot @teghsm and more! product side: - travel portal: 5% bitcoin back on flights, hotels and cars, no monthly cap. lives in the app, built with rocket travel - secured card: post $500 to $5,000 in usdc as collateral to get approved with no credit history, and you still earn rewards on the deposit - web onboarding - card management improvements - many more improvements and more coming :) feels like we were an army of 100, but we're ~9 ICs in the card team with @vladberteanu. the company makes it easy to partner across orgs and ship together, which is most of why any of this works.
BREAKING: Anthropic CEO says open-source AI is moving down a “very dangerous path.”
Coinbase is the financial account for the intelligence age.
How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching. Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting to open weight models like GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task. 91% of our employees were never hitting their usage caps, so instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models, so they can check each other's work. Better Routing – In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing. For instance, you may want a frontier model for planning, but not for execution where they can be overkill. Ultimately, humans shouldn't be choosing models - AI can automate this task. Better Caching – Cache misses are the easiest way to drive your cost up. All of our requests are cache aware, so we’re reusing a warm cache wherever possible. For example, our cache hit rate went from 5% → 60% in LibreChat once properly implemented. Keep Context Lean – Start fresh sessions when switching tasks. Scope file context narrowly. Disconnect unused tools. Don't just compact. The goal isn't fewer tokens used, it's fewer tokens wasted. Better Visibility – Our engineers can use as many tokens as they want, from whatever model they want, but we’ve made usage visible – and the more you spend on AI, the more impact we expect. The goal isn't to suppress usage. It's to build the infrastructure that makes exponential growth sustainable. Putting this into practice has cut our AI spend nearly in half, while our token usage continues to grow.
talk to a friend's codex
Coinbase is also using @linear as the backbone of the operations. Their internal agent can action from there and choose the models as needed
How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching. Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting to open weight models like GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task. 91% of our employees were never hitting their usage caps, so instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models, so they can check each other's work. Better Routing – In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing. For instance, you may want a frontier model for planning, but not for execution where they can be overkill. Ultimately, humans shouldn't be choosing models - AI can automate this task. Better Caching – Cache misses are the easiest way to drive your cost up. All of our requests are cache aware, so we’re reusing a warm cache wherever possible. For example, our cache hit rate went from 5% → 60% in LibreChat once properly implemented. Keep Context Lean – Start fresh sessions when switching tasks. Scope file context narrowly. Disconnect unused tools. Don't just compact. The goal isn't fewer tokens used, it's fewer tokens wasted. Better Visibility – Our engineers can use as many tokens as they want, from whatever model they want, but we’ve made usage visible – and the more you spend on AI, the more impact we expect. The goal isn't to suppress usage. It's to build the infrastructure that makes exponential growth sustainable. Putting this into practice has cut our AI spend nearly in half, while our token usage continues to grow.
