I Need Coffee - Episode 208 - Weekly BC Review! (April 17, 2026)

1. Weekend – World of Coffee Conference (San Diego)

Attended with friend Gabe from San Salvador (his coffee business ranked #69 in North & Central America – huge achievement).
Massive event: thousands of hardcore coffee nerds, huge convention center, top roasters (Onyx, Weber, Probat, etc.).
Highlights: premium packaging, wild processing methods (double fermented, anaerobic, nitrogen infused, co-fermented lychee, etc.), latte art competition, high-end light roasts.
Bought premium specialty coffee (~$30 per small box).

2. Customer AI Stories & Warnings

Customer used ChatGPT to justify reusing check numbers in BC → team had to intervene with real risks.
AI responses should be taken with a grain of salt — especially for ERP/accounting decisions.

3. Legacy PTE Nightmares (Inherited Lot Tracking)

11-year-old PTE (from ~2015, migrated from C/NAV) that bypasses all native BC lot tracking.
Automatically assigns lot numbers to everything while keeping items as non-tracked.
Causes quantity discrepancies and breaks standard inventory adjustments.
Decision: Do not try to fully fix/migrate — just patch bugs. Classic “shadow system” horror story.

4. Hardware Upgrades

Ordered Mac Mini (M4 base) for heavier AI workloads (replacing eBay tiny PCs).
Used OpenClaw + custom “lobster shopper” script to monitor Apple refurbished store every 30 mins.
New Human Scale G5 keyboard tray (Rolls-Royce of trays) — tilt/height adjustable, very happy.

5. AI Tooling & Vibe Coding Insights

Stefan Moron’s AL Symbols/Dependency MCP + Jeremy’s BC Code Intelligence MCP installed but inconsistent tool usage in agents.
Improving AI-generated user guides/documentation — Joselyn’s brutal (and valuable) feedback led to much better prompts.
New strict grounding rules: no hallucination, tie everything to actual code/pages/fields.

Corto vs Pandoc: Corto is mostly a wrapper; sticking with custom Pandoc + Word template workflow.
AI is terrible at time/duration estimates — completely hallucinates. Added custom hooks for real timestamps.
Claude Code + Tmux on Linux for running multiple agents efficiently.

6. Major AI Disruption Example (Survival Mode)

Friend’s company: 2.5-year, multi-million dollar custom .NET project with 8 developers.
Customer vibe-coded the entire system in 12 days (including missing features + improvements).
Result: Laying off 6 developers, major contract renegotiation.
Steve’s warning: Everyone should be in survival mode in 2026. The world is changing extremely fast.

7. Microsoft Account & Access Frustrations

Microsoft login dialog hell (no account specified) → accidentally created personal account conflict with M365 business account.
VPN/Conditional Access blocks for international team members (geolocation + other policies).

8. Open Source Contribution Fun

Helped Stefan with ALRunner project (super-fast AL unit testing, no Docker/container).
Ran multiple Claude agents via Tmux on Linux box → 100+ commits/PRs.
Cleared 1,100+ issues with several people contributing agents.

9. Model & Tooling Updates

Opus 4.7 released — even better, but very expensive on GitHub Copilot (7.5x → 15x multiplier).
Local models (LM Studio, Qwen, etc.) tested on RTX 3080 → disappointing for BC work.
Overlay VPN experiments (UniFi Teleport = garbage, Tailscale = okay but slow).

10. Team & Other Updates

Team studying MB-800, working on Document Capture, PMP, weather cam project.
Upcoming events: Directions NA, Dynamics Con, Dynamics Minds, Vibe 2.0 Partner Conference (Provo, UT – Sept 21-23).