AIMM Session — December 5, 2025: Iteration Compression and the Medical Intelligence Agent

“We never go from the idea to the finished product. We always go from idea to half-assed product to a slightly better product, to ‘yeah, I can live with that,’ to ‘oh my god, let me get some feedback,’ to ‘this is ready to be published.’ We might as well compress the cycle so it doesn’t take 2 weeks — it takes 2 days.” — Lou

30-Second Summary

The iteration compression thesis — AI’s real value isn’t replacing what you do, it’s collapsing the number of cycles between “blank page” and “ready to ship”
Enterprise adoption has a hidden gap — Leaders use AI privately; their teams don’t
Elizabeth’s medical agent needs SQL, not RAG — A shareable health-tracking agent surfaced a key architectural lesson about when vector search is the wrong tool
GEO schema: Lou’s ICH-to-FAQ tool is live — A working Streamlit app converts your Ideal Client Handbook into schema.org markup designed to get you cited by generative engines
API costs hit Dirk like a wall — $5 burned through in a single session sparked a useful pricing breakdown across Claude models

1. The Iteration Compression Thesis

AI doesn’t replace the creative process — it shrinks the loops inside it. Everything you create goes through the same rinse-lather-repeat cycle of research, draft, revise, test, refine, publish. That cycle used to take weeks. Now it can take days — not because AI writes the final product, but because it eliminates the dead time between iterations.

The research step alone tells the story. Before AI, you Googled, read papers, extracted what mattered, moved it into your workspace. That’s grunt work. The only valuable part was deciding what’s relevant and how to interpret it. AI handles the grunt work; you keep the judgment.

For people paralyzed by the pace of change, Lou offered a practical reframe. Stop trying to become an AI ninja. Instead, look at your content creation or service delivery process, find the rinse-lather-repeat loops, and ask which of those cycles AI could compress. Research? First drafts? Brainstorming? Feedback simulation? Pick one. Compress it. Then compound from there.

Lou’s metaphor: “You don’t have to start on the Autobahn. Take the countryside road. Then work your way up.”

What This Means for You: Map your content creation process as a cycle: Research → Ideate → Draft → Revise → Test → Publish. Pick the single step where you spend the most time doing work AI could handle (usually research or first-draft generation). Hand that step to an LLM for one week. Measure whether your cycle time actually compresses. If it does, you’ve found your on-ramp.

2. The Enterprise Adoption Gap: Leaders Use AI — Teams Don’t

Bally surfaced an observation: leaders are using AI and are “quite competent” with it — but their teams aren’t. The leaders feel confident enough to use the tools, but not confident enough to share that they’re using them.

Kasimir cited research confirming the gap — and a solution. Successful enterprise AI adoption requires both top-down and bottom-up movement. Top-down gives you enterprise-level harmonization: unified databases, security policies, sanctioned tools. Bottom-up gives you the innovation: people in the trenches discovering what AI can actually do for their specific workflows. Mandate-only approaches kill adoption because people disengage when they can’t explore.

Lou extended the analysis into the cultural fear layer. Employees see a rational threat: if I make AI work too well, I’ve automated myself out of a job. That produces negative compliance — people technically following the policy while quietly undermining adoption.

Hot Take: The leaders-use-it-but-won’t-say-so dynamic isn’t just an adoption problem — it’s a market signal. Every executive secretly using ChatGPT to write strategy memos is a future buyer of the consultant who can do it better, faster, and with domain expertise baked in.

3. Building a Medical Intelligence Agent — and Why RAG Is the Wrong Tool

Elizabeth has a genuinely advanced use case: a family member is critically ill. The doctors aren’t great at explaining results. Her family wants real-time access to an AI that can analyze months of blood tests, track trends, correlate values across time, cross-reference with medical research, and surface treatment insights.

The group workshopped the architecture live. Lou’s key insight shifted the entire conversation. This isn’t a language problem — it’s a data problem. Blood test values fluctuating over time, correlations between metrics, trend analysis across months: that’s SQL territory, not vector search.

RAG retrieves semantically similar text. SQL filters, computes, and aggregates numerical data.

Lou’s recommendation: store the blood work in a Postgres database, use an LLM to write SQL queries against it, then pass the retrieved data to the LLM for inference and visualization with Plotly.

The vibe-coding angle made it accessible. Lou told Elizabeth she could build this as a Streamlit app — a single Python file with a simple interface her non-technical family could use. “I find it easier to vibe code a simple app than to do a spreadsheet.”

Resources:

Streamlit (streamlit.io) — Open-source Python framework for building data apps in minutes
Plotly (plotly.com/python/) — Interactive graphing library for time-series data correlation

4. GEO Schema: The ICH-to-FAQ Tool Is Live

Lou’s been building since the last session — and the first working version is already deployed. The tool takes your Ideal Client Handbook as input and generates a complete schema.org FAQ structure designed to get you into Google’s Knowledge Graph and cited by generative engines.

The causal arc is the architecture. The system decomposes your ICH into root psychological elements, then maps them through a chain: beliefs → values → fears and aspirations → obstacles → solutions → outcomes. At each node, the schema creates entities that generative engines can follow. The end of the chain points to you as the authoritative source for that outcome.

Ken wants the God Protocol. Lou wants the minimum effective dose. His approach: ship what works, test it live, iterate from real data.

Hot Take: First-mover advantage in GEO is real and time-limited. The window to claim your domain within the knowledge graph, before competitors figure this out, is measured in months, not years.

5. API Costs: The $5 Wake-Up Call

Dirk burned through $5 in a single Typing Minds session discussing “maybe 20 pages” of documents. The group diagnosed the likely cause: running on a high-end model without realizing it.

Lou broke down the pricing tiers:

Claude Opus 4.5: $5/million tokens in, $15/million tokens out
Claude Sonnet 4.1: $2 in, $10 out
Claude Sonnet 3.7: roughly $0.30 in, $0.80 out

That’s a 15x cost difference between the top and the budget option — and for many tasks, the cheaper model is more than adequate.

Lou’s own numbers: over 6 months building a legal AI app — embedding hundreds of thousands of pages, running retrieval constantly — he spent around $100 total. The trick: using Haiku for everything that doesn’t require the heavy model.

Community Corner

Elizabeth Stief is building something genuinely useful under difficult circumstances — a medical tracking agent for a seriously ill family member. Her naming convention for Claude (“Claudette,” because in French, Claude is feminine) continues to be a detail worth noting.
Dirk is feeding his Supabase recruiting database via N8N agents that auto-process CVs, pull LinkedIn profiles, and enrich candidate records. His next step: building a Streamlit dashboard to surface and search candidates.
Kasimir reminded the group: build in Claude (best writing quality), export to Gemini (best context capacity), test in ChatGPT (best distribution). Each model has a role.
Bally grounded the session with her observation about the enterprise adoption gap — an insight that reframed how the group thinks about their own positioning as AI consultants.

Try This Before Next Session

Run the Streamlit test. If you’ve never vibe-coded anything, this is your 15-minute on-ramp:

Open Claude (or ChatGPT) and say: “I want to set up a Python development environment on my Mac/PC. I’m a beginner. Give me one step at a time — just the commands I copy and paste into my terminal.”
Follow the steps (usually 5-7 commands)
Then say: “Now create me a simple Streamlit app that takes a text input, sends it to an LLM, and displays the response.”
Run it locally.

You’re not shipping a product. You’re proving to yourself that the gap between “I have an idea” and “I have a working interface” is smaller than you think.