RecapFlow : November 3rd Coaching call analysis

📝 SUMMARY

This week's call was a small but packed session (Patrick Chouinard, Ty Wells and Juan Torres, with Shah Martinez joining later), with lower attendance due to a time change and the midterm. The group covered a lot of ground: tooling and agent setups (Ty's move to Duet as a "master agent controller", Patrick's continued use of T3 Code), Patrick's "AOE API meter" dashboard for tracking API spend, quotas and resets across providers, and Juan's WaveSpeed walkthrough with per-model tracking questions and a Minimax H3 360 Orbit LoRA demo, feeding his planned photo-to-3D-style-video pipeline for AI Boot Studio. Patrick shared his training-material automation and explained the enterprise constraints around resellers like WaveSpeed and fal.ai (hosting location, no data-retention rules, no responsibility insurance), plus his liability-driven rule that suppliers must pass full evaluation, currently pointing him to the hyperscalers, with tentative picks of an Azure OpenAI image model, Google's Gemini "Omni" for avatar video (not yet GA, so VEO3 for now) and 11Lab for voice. Ty demoed CallSitter, his hold-waiting callback app, and Patrick showed his rebuilt "Community Brain" meeting-recap app using LensDB, with Juan giving a job-search update and Biggi Fraley praising both demos in chat.

💡 KEY INSIGHTS

Ty switched from T3 Code and CMUX to Duet as a "master agent controller". It worked for him because his infrastructure was already in place: he only had to expose two MCPs, one for incoming signals and one for actions that run other agents. He keeps a Claude remote-control session as a fallback.

Ty moved to a "fleet" system built on GitHub Issues that manages multiple subscriptions, which let Duet fill his mobile gap. Patrick still likes T3, in particular its mobile remote features.

Patrick's AOE API meter aggregates monthly spend, next payment, prepaid balance, next reset and alerts across providers, and projects when an allowance will run out. Examples: Tavily, expected to run out on Nov 30 with a Dec 1 reset; GitHub Actions minutes, projected by a simple estimate from real usage against an even-pace line.

The meter can show per-key and per-model breakdowns only where the provider's API supports it. OpenRouter gives rich per-key and per-model data. Tavily does not expose which key made a call.

Patrick creates a separate OpenRouter key per project, each with its own renewal period and cap. Most use non-Claude models (GLM variants) because OpenRouter marks up Claude.

New providers can be added via YAML without coding. GitHub needed custom code. Patrick tells his agent ("Dots") to monitor the data and alert him, and uses the dashboard to investigate. Ty prefers an agent that says "yay or nay" over dashboards he has to check himself.

Juan said WaveSpeed's dashboard and (on his silver account) API give per-model accountability. Patrick had found only a single universal key with little control on his personal account. The difference between account tiers is speculative.

Juan noted 11labs is attempting WaveSpeed-like node workflows, with fewer models, but lets you create an API from a workflow. WaveSpeed has little documentation on this, so he suggests building your own API chain.

Juan said Minimax H3 is open source, supports LoRAs, and is "as good or even better than" C-Dance (his claim). With the Minimax H3 360 Orbit LoRA, one photo yields a 3D-style turnaround of a person or object.

Juan's planned pipeline for AI Boot Studio: photo, background removal, LoRA video generation, upscale, then delivery to the guest via QR code. WaveSpeed API calls are primitive (image-to-video, text-to-video, etc.) and a saved desktop workflow can't be called as a whole, so he is researching how to chain them. He also wants to store model settings in a database.

Ty suggested upsell ideas: animate the person's 3D model (backflips, beach, ocean). Juan said both whimsical and elegant variants would be needed.

Patrick uses WaveSpeed mostly as an "open router" for models, called sequentially. His media pipeline is Scott's, which he extended with FinOps tracking and supplier flexibility (Higgsfield, 11 Labs, WaveSpeed, OpenRouter, possibly fal.ai).

Patrick's focus is script and training-material generation. The pipeline uses computer use to follow product documentation, take screenshots, build animated graphics with an avatar, script and text-to-speech. He wants it eventually to watch repos and changelogs to update training automatically.

Ty argued that learning now happens by conversing with LLMs rather than one-way video. Patrick agreed in principle but said corporate users aren't ready (some are unsure what a prompt is, or confuse Claude inside Copilot with Claude itself).

Patrick's Claude Code training for his client is a Claude harness/project containing only what differs from Anthropic's material, so learners can ask questions, go on tangents and resume. He said this suits developers, not executive assistants. He also said a four-month course is obsolete before release, so training must be generated in minutes to hours.

Patrick has an automated process that publishes VS Code / GitHub Copilot release notes contextualised to company governance. It writes a Confluence page, then a summarised team message, plus a bi-weekly 30-minute deep-dive meeting. He replicated it for Claude Desktop, Claude Code and other software, and is working with the dev team to use it for internal-software training. Juan suggested an AI-generated video of Patrick summarising the bi-weekly meetings.

Patrick said he can use WaveSpeed, fal.ai and similar services personally but not in the enterprise. Reasons: they resell models hosted elsewhere (mostly Chinese models hosted and run in China), have no data-retention rules and no responsibility insurance. WaveSpeed is responsible only for "the first $100" of problems (statement cut off in the source).

Patrick: in a corporate/financial setting every model and supplier must pass a full evaluation covering data location and liability. The answer to "who is responsible if something goes wrong" can never be "us", so liability is pushed to a provider willing to take it. Currently only the hyperscalers (Google, AWS, Azure) are ready to do that.

Patrick's reading of provider indemnification: if a generated deepfake (his example was a CEO saying something scandalous) leads to a lawsuit, a provider would cover only a nominal amount and the rest is on the customer. He says he cannot sell that corporate-wise.

Patrick's tentative enterprise picks:

Image: the Azure OpenAI image model (he named it "GPT image 2.5", "if that's correct").

Video avatar: Google's Gemini model "Omni", which he calls decent but not C-Dense (as transcribed) quality.

Voice: 11Lab.

Enterprise policy forbids pre-release models, so a model must be fully GA and supported. Omni is not GA yet, so he has to work with VO3. He thinks 3.1 might be usable, but Omni is not.

Patrick uses Juan's pipeline to learn what the mechanics can do, then swaps in models that pass security and liability review.

On resellers, Patrick's research suggests WaveSpeed is much smaller than expected and is a reseller. Juan agreed they provide model access rather than creating models, though they run inference engines for open-source models (Juan mentioned Verda). Patrick says Runway ranks its own model third behind Seedance and Kling, and that ElevenLabs' video generation resells Chinese models (C-Dance/Kling). Juan said ElevenLabs doesn't allow C-Dance because it is Chinese, then agreed the API reseller path exists. Patrick uses ElevenLabs for voice because that is their own model.

If governance blocks 11Labs, Patrick may have to go full Google. He said Google has decent voice models (not 11 Lab quality) and a music model, Lyra, a Suno competitor that is "supposedly" not Suno grade but pretty decent. Juan said he learned of it this week.

CallSitter (Ty): the app places the call to the business, helps the user get identified, then waits on hold. Once the user has authenticated and is on hold, the user hangs up (star, star). The app watches hold progress and calls the user back when a human is about to come on. It does not record or keep data.

CallSitter pricing as Ty described it: 10 cents a minute, with a minimum 15-minute hold. A hold under 15 minutes costs nothing. A paywall link is sent when the callback arrives, with a hold on the card. Charges start at $2.49 at 15 minutes and max out at $8.99.

Callback handling: users are told to save the callback number via a vCard link so screened calls come through. If the user misses the callback, the agent retries a couple of times and the user still pays.

Ty's market data in the demo (U.S. only) were not accurate. He suspected the wait-time numbers came from GetHuman and said he needs to check them.

Patrick's concern: companies bank on callers dropping off, so widespread use of such a tool could affect wait times. Ty's response: they will need to fix the root cause. Patrick also warned that government lines may hang up within half a second if no live person answers.

Patrick's Community Brain pipeline:

He scrapped the N8N workflow and rebuilt it as a web app, since the N8N workflow was generated by Cloud Code anyway.

It holds every meeting since the group began.

For any meeting it produces the next week's invite and community post message, raw and rendered, with copy and download options.

It shows per-step models, per-step run progress, per-chunk progress and live activity.

A run took about 15 minutes. Patrick says he sometimes posts the recap within 15–20 minutes of the call ending.

Long outputs are chunked when they would exceed about 32k output tokens.

Models come from OpenRouter, with per-model details shown (example: 1M context, 65,000 output, 25 cents in / $1.50 out).

The indexing models for LensDB trigger a warning if changed and should not be changed mid-run. The embedding is fixed (gnomic, run locally).

LensDB is a portable, self-supporting vector DB file, chosen over PG vector so users don't need to run a server.

Open-source status: the pipeline app is not open source yet, but Patrick could open-source it. The compiled LensDB vector DB is open source in a distributed project that needs your own front end (he proposes OpenWeb UI), though he is building his own chatbot on top.

Juan started systematizing his job application process, targeting full-stack or forward-deployed engineer roles at diffusion-model providers. He got a fast rejection, which he thinks came from screening, since it was only three days after applying. He will keep pursuing clients through his AirBoot Studio project.

Biggi Fraley (chat) said too many companies give no option but holding or trying back, and praised Ty's app. She also said Patrick's Community Brain "has come a long way".

❓ KEY Q&A

Q (Ty Wells): Does the meter show which project the usage is going to?

A (Patrick Chouinard): Not yet. It depends on the platform's API supporting it; for Tavily he can only see call counts, not which key made the call.

Q (Ty Wells): Is OpenRouter's Claude usage cheaper than Anthropic's API, or do they mark it up?

A (Patrick Chouinard): Agreed there is a markup; that is why most of his OpenRouter keys use other models. Only one key, community brain, uses a Claude model.

Q (Juan Torres): Per project, can you see a breakdown of each model used?

A (Patrick Chouinard): Yes per key and per model, but the key hadn't run yet in this period; the dev key has only one model.

Q (Juan Torres): Are you tracking the models you use for WaveSpeed?

A (Patrick Chouinard): No; he has only a personal account and the billing API doesn't give it. Juan said he gets per-model accountability on his silver account. Patrick described his own account only as "pure exact" (unclear in the transcript) and will look into it.

Q (Patrick Chouinard): Once a workflow is set up and saved in WaveSpeed, can you call it programmatically?

A (Juan Torres / Ty Wells): Juan said he plans to use the WaveSpeed API to call everything programmatically. Ty said a workflow can't be called as a whole, but API calls can reproduce the same chain.

Q (Ty Wells): How are other companies handling training with AI: are you ahead?

A (Patrick Chouinard): Not necessarily ahead; everyone is redefining training, and long course-creation cycles are no longer viable.

Q (Ty Wells): Not even 3.1 (for the enterprise)?

A (Patrick Chouinard): He thinks 3.1 might be possible, but Omni is not GA and so cannot be used.

Q (Patrick Chouinard): Once the person answers at the other end (e.g., a government line), what happens in CallSitter?

A (Ty Wells): The agent tells the answerer to hold for the user, who is connected. Ty acknowledged Patrick's point that some government lines would hang up immediately, and said the call window might need to be tuned dynamically from learned call-volume data.

Q (Biggi Fraley, in chat; relayed by Patrick Chouinard on the call): How long does patching you in take, from the agent hearing a human to your phone ringing?

A (Ty Wells): Seconds. It is a data call, so the callback comes right away and the user just has to answer.

Q (Juan Torres): Is Community Brain open sourced?

A (Patrick Chouinard): Not this one yet, but he could open-source it; nothing in it is confidential. The compiled LensDB/vector DB is open source.

Q (Juan Torres): Does it account for token usage on each iteration?

A (Patrick Chouinard): He answered in terms of output limits: if a step would exceed about 32k output tokens, it is cut into chunks. He also shows each model's context, output limits and pricing in the settings.

🛠️ TOOLS AND CONCEPTS MENTIONED

Duet – Ty's mobile-capable master agent controller, connected via two MCPs.

T3 Code – Patrick's current platform; praised for its mobile remote features.

CMUX – Ty shut it down after adopting Duet.

Claude Code / Claude remote control – Ty's fallback session; Patrick uses Claude Code for pipeline and training work.

Cloud Code (as transcribed in part 2) – generated the original N8N workflow.

Proxmox – Patrick installs it on every new machine as a cluster compute node.

GitHub Issues / GitHub Actions – Ty's fleet system uses issues; Actions usage drives costs.

AOE API meter – Patrick's custom dashboard tracking API spend, quotas and resets.

Tavily – search API tracked in the meter (spelled "Tattly" and "Tavoli" in the transcript).

Firecrawl – API tracked in the meter.

OpenRouter – per-project keys, rich usage data, source of per-step models in Community Brain; marks up Claude.

GLM 5.3 flash and GLM 5.2 – models used on Patrick's OpenRouter keys.

Codex – tracked alongside Claude usage.

Anthropic API – subscription and credit tracking; no five-hour counter on Ultra.

Hermes – Patrick wants to add it as a collector; it handles research.

Dots – Patrick's agent monitoring the meter's data.

OpenAI Direct, Perplexity, Onshow – other tracked APIs (the transcript is unclear on "Onshow").

ElevenLabs / 11Labs / 11Lab / "11 Labs" – voice provider (spellings as supplied); Ty also uses it for voice cloning.

WaveSpeed (and its desktop app) – media model provider and reseller; Juan's workflow and LoRA tool.

Verda – mentioned by Juan as an inference engine used by WaveSpeed.

Minimax H3 – open-source model per Juan; supports LoRAs.

Minimax H3 360 Orbit LoRA – turns one image into a 3D-style turnaround.

Hugging Face – source of the LoRA URL.

Seedance / Seedream (transcribed "Seedense, Seedream"), C-Dance / C-Dense, Kling – video models referenced; spellings as transcribed.

Runway – video provider that ranks Seedance and Kling ahead of its own model, per Patrick.

Higgsfield ("Higgs field") – provider in Scott's pipeline.

fal.ai – possible future supplier; usable personally but not in Patrick's enterprise.

Azure OpenAI image model – Patrick's likely enterprise image provider; he called it "GPT image 2.5".

Google Gemini "Omni" – candidate video avatar model; not yet GA.

VO3 – Google video model currently usable in Patrick's setting.

Google 3.1 – version mentioned as possibly usable; no further detail.

Google Lyra – Google's music model, a Suno competitor.

Suno – music generation reference point.

Confluence – receives the automated release-note pages.

Azure DevOps – company's tool in place of GitHub.

VS Code and GitHub Copilot – sources of the automated release-note process.

Claude Desktop – the release-note method was replicated for it.

GitHub Copilot – reference to users who confuse it with Claude.

CallSitter (also "call center.app" / "Call sitter.app" in the transcript) – Ty's app for waiting on hold on the user's behalf; https://callsitter.app/

GetHuman – suspected source of the wait-time data in Ty's app.

TaskRabbit – analogy for CallSitter ("digital TaskRabbit").

Community Brain – Patrick's meeting-summary and invite pipeline web app.

N8N – earlier workflow tool, now scrapped.

LensDB – portable vector DB file used by Community Brain.

gnomic – fixed local embedding model.

PG vector – rejected alternative because it needs a server.

OpenWeb UI – suggested front end for the open-source vector DB.

Fathom – meeting recorder; its link is used to fetch transcripts in Community Brain.

Zoom – Zoom chat file to be auto-grabbed via Mac Mini.

Mac Mini – Patrick's machine for upcoming auto-fill of the Zoom chat.

AirBoot Studio / AI Boot Studio (both spellings as supplied) – Juan's project for attracting clients.

Stacks – tool Ty mentioned using with 11 Labs (context unclear).

read.ai and Fireflies.ai notetakers – added to the meeting by attendees (per chat log).

📎 SHARED RESOURCES

https://callsitter.app/ — CallSitter, shared by Ty Wells in chat.

🔄 FOLLOW-UPS WORTH EXPLORING

Open questions, follow-ups and commitments:

Patrick Chouinard will look into WaveSpeed's API/billing options for per-model tracking.

Patrick Chouinard intends to add Hermes as a collector in the meter and reload his machines.

Juan Torres plans to build an API-based pipeline (photo, background removal, LoRA 360 video, upscale, QR delivery) for AI Boot Studio.

Juan Torres plans to store model settings in a database; he intends to send Patrick the LoRA.

Patrick Chouinard has already submitted his AI-generated media/training proposal to his boss for AI committee approval; it must be reviewed.

Patrick Chouinard is working with the dev team to apply the release-note-to-training method to internal software.

Patrick Chouinard will likely go with the Azure OpenAI image model, Google's Gemini Omni for video avatars and 11Lab for voice, with a possible fallback to full Google if governance blocks 11Lab.

Patrick Chouinard will work with VO3 for now, since Omni is not GA.

Patrick Chouinard plans an upgrade to Community Brain after the day's processing, with auto-fill of the Fathom link and Zoom chat and a built-in chatbot, expected from next week.

Ty Wells will check the accuracy of the wait-time numbers in CallSitter.

Juan Torres will continue working on his application process and trying to get clients through AirBoot Studio.

0 comments

RecapFlow : November 3rd Coaching call analysis