Who was in the room

Session 2 drew a healthy cross-section of the network — from people just getting started to seasoned practitioners with deep production experience. The mix was a strength: beginners felt safe asking foundational questions, and the advanced voices added texture without talking over anyone.

Co-leads: Ivy Marguerite set the agenda and steered the Q&A; Sarah Gudeman (BranchPattern, Omaha) opened with the group's origin story and kept the energy moving. Featured presenter: Karan Khanna, CTO at EyeQ Monitoring, delivered a pragmatic ~20-minute landscape talk — comfortable going deep on tokens, edge models, and agent orchestration, but consistently translating back to everyday CRE use cases.

Where the puck is going

• Models will keep improving for at least the next few years — no plateau yet.
• Specialized vertical models are emerging, including real-estate-focused ones.
• Token / inference costs are dropping fast and will continue to.
• Edge models (running on laptop/phone) will become more prevalent — less cloud dependency, lower cost.
• Agents are arriving — most mature in software engineering today, but spreading fast.
• Any frontier model handles CRE use cases well. Copilot still lags but may catch up via the Microsoft–Anthropic deal.

How to choose a tool (in priority order)

• Company policy — if mandated, use it.
• Security — use whatever IT has blessed.
• Enterprise version over public (your data isn't used for training).
• Paid over free (better context retention, faster, ~$20–30/mo).
• Context window size for large documents.
• Integration with your productivity suite (Google Workspace vs. M365).

Gotchas & good practices

• Don't violate company policy.
• Watch token costs on long agentic tasks (transactional use is fine).
• Always do a final human review.
• Use multiple models to learn their quirks; experiment in personal domains, then translate to work.

Q&A highlights

• Token costs: a few hundred tokens per chat, dropping fast — a one-hour Claude agent run might cost ~$1 today.
• Michelle on Claude: argued it's the most "humanized" model, especially for voice-matching and personality work.
• Claude "skills": build hierarchical skills assigned to parallel agents with an orchestrator collating results. Meta-prompt Claude to teach you how to build them.
• Meeting notes: Fireflies for API-friendly auto-sharing; also the Bee wearable and Whisper for dictation.
• Recording legality: one-party vs. two-party consent varies by state — disclose and offer to share the transcript.
• Security: don't give a model PII or financial info without an enterprise version. The real risk is shadow AI when firms over-restrict and employees route around policy. Aim for smart-but-not-restrictive governance.
• Redaction: use Adobe's redaction tool on sensitive PDFs before uploading.

Closing challenge

Between now and next session, each member tries one thing — a workflow to speed up, a time sink to automate, or a new way to get information faster.

Recurring logistics: 3rd Friday of every month, 10:00 AM ET. Email Ivy at imarguerite@eyeqmonitoring.com to be added, or register for the webinar.