RecapFlow : October 20th Coaching call analysis

📝 SUMMARY

This week's call packed a lot in: it opened with casual hardware talk that turned into a bigger observation about AI widening the gap between tech workers who adapt and those who don't, then moved through a wide range of member updates — Patrick Chouinard's company-wide Claude rollout and personal tooling projects, Juan Torres's photo booth business, Andrew Nanton's document-review and decision-model work, Paul Miller's productivity gains from pruning Claude skills, Ryan C's custom Android remote-management software, and Daniel Zivkovic's pivot toward AI governance work. The second half went deeper into technical and business debates, including whether self-hosted inference (VLLM/TensorRT LLM) can beat commercial APIs (a discussion where attendees largely converged on APIs rarely being beaten without a governance or scale justification), Canadian AI governance trends, advice on entering financial-industry contracting, project demos from Shakur Abdullah, a comparison of a code security/review tool against Greptile and CodeRabbit, and a closing segment on AI video-generation tooling for corporate training. Whether you're interested in enterprise AI adoption, infrastructure choices, governance, or hands-on tooling, there was something here — and the recording captures it all.

💡 KEY INSIGHTS

Patrick Chouinard: people aren't replaced by AI directly, but by colleagues who adapted faster — those not keeping up are effectively "signing their name on the list" for future layoffs.

Patrick Chouinard: just keeping up with the pace of AI field evolution is now filling people's schedules.

Paul Miller suggested the surge of interest in his client's AI training courses may be partly driven by fear of job irrelevance rather than genuine learning interest; Patrick Chouinard agreed, noting a split between people building capability and people attending out of fear.

Patrick Chouinard: data providers (e.g., Bloomberg) weren't ready for the AI "tidal wave" — restrictive data-usage contracts must now be rethought for MCP-based direct AI access.

Patrick Chouinard: significant career opportunity exists in enterprise AI integration/connector work, since "best practices" go stale within a week and nobody has fully solved the problem.

Patrick Chouinard: building a marketplace/pipeline for AI skills/plugins is harder than expected — deduplication/evaluation of natural-language artifacts has no existing framework, unlike structured datasets.

Paul Miller: running a "skills removal and review" via Claude Doctor produced a dramatic quality improvement within 24 hours, especially with "Opus 5.5," since too many loaded skills can undermine output quality.

Andrew Nanton: there's a wide spread between people fully engaged with AI and people who don't know where to start; the only wrong approach is ignoring it (with Patrick Chouinard's caveat that even near-retirement workers may not be exempt).

Andrew Nanton: decision-model APIs (comparing Microsoft's "Decisions One" and another referred to as "Luna") are impressive in result quality even if not fastest/cheapest; he floated domain-specific fine-tuning of smaller decision models given their similarity to BERT-style models.

Patrick Chouinard: small decision models could become valuable for edge/on-device decision-making at scale.

Andrew Nanton: "boring, voluminous, and recurring" tasks (document review, tagging, vector search, timelines) are the best current fit for AI in his forensic work — not the most exotic use cases but the most consistently useful.

Daniel Zivkovic: Canada has no distinct "AI governance" role yet — it sits under data governance; he expects this to change, drawing a parallel to the Basel Capital Accord era as a "fishing trip" for consulting opportunity.

Daniel Zivkovic: OSFI has a compliance deadline for Canadian banks next May, which he's preparing materials for.

Patrick Chouinard: integration and governance will see the most work growth over the next year; opportunity is gated more by people's adaptability than AI's technical pace.

Patrick Chouinard: good AI use-case criteria are boring, high-volume, and repeatable tasks.

Daniel Zivkovic: financial institutions are cautious about production AI use cases — most effort goes into approval processes, so intake/triaging is a safer entry point.

Juan Torres: some inference providers use "Verda" as a specialized GPU-only cloud provider, differing from full EC2-style hosting.

Daniel Zivkovic shared a past-project anecdote where executives bought expensive GPU hardware mainly to "justify the project," but results couldn't match commercial cloud API quality; running large open-source models locally (e.g., "OSS 120 billion") can still yield worse answers than a hosted model like GPT-4o.

Juan Torres: for his accounting-firm entity-recognition use case, large APIs (e.g., GPT 5.1) significantly outperform smaller self-hosted models (8B on A10 GPUs), making self-hosted inference not worth it there.

Paul Miller recommends benchmarking with top-tier models first, then testing cheaper models (via Open Router/Bedrock) against synthetic data to find the best price/quality tradeoff, possibly with a router/hybrid approach.

Daniel Zivkovic: such benchmarking is a good neutral way to show clients whether building their own inference engine is worth it.

Daniel Zivkovic: AWS and Google are "opinionated" providers while Azure is comparatively generic, succeeding mainly via its OpenAI partnership; Google Cloud is easiest to set up AI with.

Patrick Chouinard: uses Open Router personally, but for business/financial work uses AWS Bedrock for governance/data-residency reasons (Open Router resells inference without full visibility into data destination); still uses Azure for some models (e.g., Cohere) unavailable on Bedrock.

Daniel Zivkovic: banks favor Databricks partly for its governance pitch — Unity Catalog's logging/role-based control suits regulated industries, "even though nobody uses it well."

Daniel Zivkovic: breaking into financial-industry contracting as an independent typically requires going through reseller/consulting firms; mid-sized (100–2,000 person), often India-based companies are a "sweet spot" since they're more open to contractors and don't require in-office presence.

Patrick Chouinard: has stayed with the same mid-sized consulting firm for 17 years, valuing being known within the hierarchy over being "a number" at a large firm.

Shakur Abdullah: his multi-model "horse game" experiment found Opus 5.5 the standout "boss" model for orchestrating others; he hasn't found a use case where small models outperform.

Shakur Abdullah: working on Remold CRM, wanting integration with AI agents (Dots, Muse, GrokBot) for social media; needs someone with an M365 account to test it.

Ryan C: Scott's push-triggered review tool runs security, performance, quality, docs, release, Anthropic security, and "codex adversarial" checks on every push (not just PRs); nearly $1,798 spent this month, catching hundreds of issues; uses existing Claude/OpenAI accounts rather than separate API credits.

Ryan C: it's more comprehensive than CodeRabbit/Greptile since those are PR-based, whereas Scott's tool checks every push — useful for solo devs who don't use pull requests.

Paul Miller: T3 code beta is "10x from last week," apparently coordinating/distributing work across multiple machines.

Shakur Abdullah: with new T3 code behavior, his agents work as a team across his machines with queuing based on available capacity, generating multiple PRs in parallel.

Patrick Chouinard: T3 code and Scott's tool aren't competitors — they serve different needs and can coexist.

Patrick Chouinard: uses Scott's video pipeline plus Claude Motion, Higgsfield, and ElevenLabs for professional corporate training "capsules."

Patrick Chouinard: Claude Motion is Enterprise/Teams-tier only (not even Max 20x), generates background sound (not voice), and is fully code-generated rather than bitmap-based — strong for screen-animation tutorials but less realistic than Higgsfield for other uses.

Patrick Chouinard: his team needs a new training capsule roughly every week since capsules risk obsolescence within 1–2 weeks as Anthropic releases updates, driving automation via Scott's pipeline.

Ty Wells: used ElevenLabs v4 (with emotion tags) for his AI Practitioner Summit presentation voice, through three cuts; the visual motion was not created with Claude Motion or Higgsfield.

❓ KEY Q&A

Q (Paul Miller): Are you seeing a common set of questions across training sessions that feeds into future training?

A (Patrick Chouinard): The most common question is about token limits/indexing; interest centers on new capabilities, especially connectors, since people already had Copilot.

Q (Paul Miller): Is part of what's driving people's need to attend these AI courses protection of their own jobs, not just learning a new tool?

A (Patrick Chouinard): Yes — two groups exist: genuine capability-builders and people attending out of fear of irrelevance, reinforced by a 30-day mandatory training window or license revocation.

Q (Juan Torres): Was Open Router one of your providers in your FinOps monitoring platform?

A (Patrick Chouinard): Yes, it's one of the big providers tracked, with keys per internal app to monitor usage and token exhaustion.

Q (Juan Torres): Were you considering Tailscale, and what would it cost at scale?

A (Ryan C): Tailscale wouldn't do what's needed (root-level access like TeamViewer); TeamViewer at ~2,000–2,500 screens would cost ~£50k/year, so he built his own remote management software instead, mostly for free aside from some Cloudflare costs.

Q (Juan Torres): Were you going to use Cloudflare or AWS as the connection method, and what would it cost?

A (Ryan C): Already built with Cloudflare and works well; remaining issues are around remotely updating software on the Android boxes, not the connection itself.

Q (Juan Torres): Is the remote access for troubleshooting, or why access that many devices?

A (Ryan C): It lets him troubleshoot remote screens without physically driving there, resolving ~99% of issues and saving on travel costs.

Q (Patrick Chouinard): For your decision models, have you tried running any locally?

A (Andrew Nanton): Tried one called "Lex" (name uncertain); okay but weaker than "Luna" via the Decisions API in confidence/probability output, though not as fast or cheap.

Q (Daniel Zivkovic): What forensic psychiatry problem are you solving right now?

A (Andrew Nanton): Expert witness work on cases with thousands of pages; AI work focuses on tagging, navigating, classifying, and retrieving previously-overlooked details (structured reviews, vector search, timelines).

Q (Daniel Zivkovic): So like a legal assistant?

A (Andrew Nanton): Yes, in some ways — a hybrid plain-text and semantic search tool to recall previously-seen but initially unimportant details.

Q (Juan Torres): Has anyone built their own inference engine (e.g., AWS EC2/GPU), and does anyone have experience with Verda as a GPU inference provider?

A (Daniel Zivkovic): Used VLLM on a past project driven by executives who'd bought GPU hardware to justify the purchase; results couldn't match cloud quality. Also used TensorRT LLM previously. Self-hosting large open-source models can still underperform hosted models like GPT-4o, making it worthwhile mainly when the client has budget (e.g., Canadian government AI server initiatives).

Q (Paul Miller): How much GPU performance are you looking for, and is privately hosted inference driven by need or cost savings?

A (Juan Torres): Not cost savings — for his accounting firm's entity-recognition task, a self-hosted 8B model on an A10 GPU can't match GPT 5.1's accuracy, so a custom inference engine doesn't make sense there; his question relates to job postings requiring diffusion-model inference deployment experience, where some providers use Verda.

Q (Juan Torres): With AWS Bedrock, do you just use a Unify API without setting up inference modules yourself?

A (Patrick Chouinard): Correct; Databricks model serving is another option but more niche, mainly for scenarios already in Databricks — his team primarily uses AWS.

Q (Juan Torres): Where would be the best resources to find financial-industry contract work?

A (Daniel Zivkovic): Go through resellers/consulting companies; his sweet spot is mid-sized India-based firms (not large firms with 5x markups) since they don't require in-office presence.

A (Patrick Chouinard): Agreed — financial clients won't contract directly with freelancers due to liability/insurance requirements; offered to have Juan message him directly about his firm.

Q (Shakur Abdullah): How does Scott's security/review tool compare to Greptile or CodeRabbit?

A (Ryan C): More comprehensive and functionally different — CodeRabbit/Greptile are PR-based, while Scott's tool checks every push, which matters for solo developers not using pull requests.

🛠️ TOOLS AND CONCEPTS MENTIONED

Proxmox / PBS (Proxmox Backup Server) — home backup rebuild (Patrick Chouinard).

Claude / Claude Code / Claude Doctor — enterprise Claude license rollout; Claude Doctor used to prune excess skills (ships with Claude Code via /doctor per Daniel Zivkovic; community repo shared by Juan Torres).

Opus 5.5 — model sensitive to skill overload; standout "boss" model in Shakur Abdullah's horse game test.

Claude Co-Work — merging with chat, accelerating enterprise agentic rollout.

Claude Motion — code-generated training video/motion tool, Enterprise/Teams-tier only.

MCP (Model Context Protocol) — connects AI platforms to data providers like Bloomberg.

Bloomberg — data platform example of slow AI-access adaptation.

OpenRouter / Open Router — broad model access (Patrick Chouinard personal use; recommended by Daniel Zivkovic/Paul Miller for benchmarking).

Eleven Labs / ElevenLabs (v4, emotion tags) — voice generation (Ty Wells presentation; Patrick Chouinard's training capsules).

Firecrawl, Huncho (spelling uncertain), Tavali (spelling uncertain), OpenAI API, Anthropic API — providers in Patrick Chouinard's FinOps dashboard.

GitHub (Actions/builds, push triggers) — usage tracking and Scott's tool trigger.

T3 code / T3 beta — agentic coding orchestration platform; rapidly improving multi-machine coordination.

GitIgene (name as spoken) — Patrick Chouinard's scheduled cleanup tool for T3 code projects.

Hermes — job dispatcher to T3 code.

Codex, Claude code, "anti-gravity"/"Antigravity" — agentic coding backends orchestrated via T3 code/Hermes.

Cursor — $20/month coding tool (Juan Torres), considering switch to T3.

Tailscale — considered by Ryan C, didn't meet root-access needs.

Chrome Remote Desktop / TeamViewer — reference points for remote device management (TeamViewer ~£50k/year at scale).

Cloudflare — connection layer for Ryan C's custom remote management tool.

Screenly — Ryan C's digital signage product.

OpenAI Decisions API — evaluated by Andrew Nanton.

Microsoft "Decisions One" (Azure) — evaluated by Andrew Nanton.

"Lex" (local model, name uncertain) — tried by Andrew Nanton, weaker than "Luna."

"Luna" — decision model via Decisions API, stronger cloud option per Andrew Nanton.

Gemini — used for handwritten bounding-box OCR (Andrew Nanton).

BERT-style models — architecture reference for decision-model fine-tuning.

JEV / Jev — mentioned in chat (Shakur Abdullah, Andrew Nanton) re: entity-recognition training.

VLLM — self-hosted inference engine (Daniel Zivkovic, past project).

TensorRT LLM — NVIDIA inference engine (Daniel Zivkovic, earlier project).

Verda — GPU-only cloud inference provider (verda.com).

AWS Bedrock — business inference platform (Patrick Chouinard).

Databricks model serving / Unity Catalog — niche inference hosting and governance feature.

Cohere — model available on Azure but not AWS Bedrock for Patrick Chouinard's team.

Google Cloud / AWS / Azure — compared as cloud AI providers.

Fireworks — mentioned in chat (Andrew Nanton) as "supposed to be pretty good."

Qwen 3.8 flash next — small model praised by Andrew Nanton in chat.

Astra — mentioned by Juan Torres (chat), used for 3D modeling of building blueprints.

Remold CRM — Shakur Abdullah's in-progress CRM.

Dots, Muse, GrokBot — AI agents for Remold CRM social-media integration.

M365 (Microsoft 365) — needed for Remold CRM integration testing.

Boring Funnels Library / "Boringfunnels" — marketing funnel resource Shakur Abdullah considering.

Convex, Clerk, WorkOS — auth/backend services cycled through by Shakur Abdullah.

Scott's code security/review tool — called "CC Security Review," "Tide code"/"Tidecode" in the call, and "Tightcode"/"Tight code" in chat (site: tightcode.dev); push-triggered multi-model review (security, performance, quality, docs, release, Anthropic security, codex adversarial checks); ~$1,798/month spend reported by Ryan C.

Scott's video pipeline ("video-generator") — built on FFmpeg and Remotion, used for training capsules.

FFmpeg / Remotion — underlying tools in Scott's video pipeline.

ccblackbox.app — Scott's IDE tool with mobile companion app, used daily by Ryan C.

poweryourprocess.ai — Scott's site hosting open-source and software projects.

Greptile — PR-based AI code reviewer, free for open-source projects.

CodeRabbit — PR-based code review tool compared against Scott's tool.

Higgsfield — realistic video generation service for training capsules.

superpowers-lite — tool Andrew Nanton considering moving to (chat).

Rex — tool Andrew Nanton mentioned using recently (chat), "so far so good."

Moo / Moo Print (moo.com) — business card printing service (Paul Miller).

One Stop Creative Agency (onestopcreativeagency.co.uk) — Ryan C's business card design referral.

Vercel ("Vasell" per transcript) / Netlify — deployment targets for Ryan C.

Theo — person/content referenced for T3 code coverage, not a tool itself.

📎 SHARED RESOURCES

Links and Resources

Business cards

https://onestopcreativeagency.co.uk/ — Ryan C's business card design referral.

https://www.moo.com/us/business-cards — Paul Miller's "Moo Print" business card suggestion.

Claude tools

https://github.com/millionco/claude-doctor — Juan Torres shared; Claude Doctor tool.

https://code.claude.com/docs/en/skills#find-unused-skills — Paul Miller's link for reviewing Claude's own skills.

AI governance and performance engineering

https://magmainc.ca/guides/ai-governance-canada-field-guide/ — Daniel Zivkovic's AI governance field guide for Canada.

https://github.com/cfregly/ai-performance-engineering — Daniel Zivkovic's recommended study resource.

https://www.amazon.com/Systems-Performance-Engineering-Optimizing-Algorithms/dp/B0F47689K8/ — Daniel Zivkovic's recommended "bible" for building inference engines.

GPU inference

https://verda.com/ — Juan Torres, re: Verda GPU inference provider.

Fun projects and demos

https://codemyvibe.com/lab/horse-games/ — Shakur Abdullah's "horse game" model-comparison project.

https://www.instagram.com/reel/Ddschp6yClQ/?utm_source=ig_web_copy_link&xtok=MzRlODBiNWFlZA==&srtk=MzRlODBiNWFlZA== — Juan Torres shared, re: Astra 3D modeling of building blueprints.

Code review tools

https://www.greptile.com/ — Daniel Zivkovic, AI code reviewer (free for open source).

tightcode.dev — Ryan C, Scott's code review tool site.

https://www.skool.com/ai-developer-accelerator/tight-code-ai-code-review-before-your-code-leaves-your-mac?p=1338592c — Ryan C's discount code link ($39 to $19 for first 3 months).

Tutorials and open source

https://youtu.be/D1E44MRofVg?si=8yNR564AkoPQc3MD — Paul Miller's recommended T3 tutorial video.

https://github.com/scott-rippey/video-generator — Ryan C, Scott's open-source video generator tool.

https://poweryourprocess.ai/ — Ryan C, Scott's website (open source + software projects).

https://ccblackbox.app/ — Ryan C, Scott's IDE tool with mobile companion app.

Presentations

https://drive.google.com/file/d/140APzvLX4BebWs6SsdoHgDlT3_755cZo/view — Ty Wells's AI Practitioner Summit presentation.

🔄 FOLLOW-UPS WORTH EXPLORING

Patrick Chouinard's organization to purchase additional Claude license packs again after exhausting prior batches.

Patrick Chouinard to continue developing visual documentation of the T3/Hermes/Codex/Claude orchestration architecture.

Patrick Chouinard to consider building a demonstration version of his personal FinOps platform with sensitive data hidden, for possible future sharing.

Juan Torres to deploy his photo booth setup at an unpaid expo event at Chula Vista golf course the coming Sunday.

Juan Torres to develop business cards and a landing page for the AI Boot Studio project.

Juan Torres to consider recording himself going through his operational process for future YouTube documentation.

Ryan C to continue separating his screen company into its own UK limited company, including VAT/tax setup and migrating clients to new contract terms.

Ryan C to continue troubleshooting remaining issues in his custom-built Android remote management software.

Ryan C to enter at least one of his two in-progress websites into the AWWWARDS competition once client approval is given.

Andrew Nanton and team to present their forensic psychiatry AI work group findings at an international meeting next month.

Daniel Zivkovic to consider attending a data management conference if recovered, in pursuit of pivoting toward AI governance/auditing work for financial regulators.

Daniel Zivkovic will share the governance guide he prepared for Canadian banks with the group.

Patrick Chouinard will connect with Juan Torres directly regarding potential contract opportunities through his consulting firm.

Shakur Abdullah plans to finish adding desired features to his project before running a full code review/security check, rather than reviewing incrementally.

Patrick Chouinard will review/listen to Ty Wells' AI Practitioner Summit presentation and may ask Claude to review it as well.

0 comments

RecapFlow : October 20th Coaching call analysis