Hermes Setup — Executive Overview

Four agents, four domains

Chief of Staff

Morning plan
1:1 prep
Gmail triage
Travel briefs

George Kuruvilla

Jess

Competitive intel
Post drafts
Deal alerts
News watch

Jackie

Net worth
529 / college
Portfolio cron
News alerts

Amit

Pre-med roadmap
Scholarships
Summer programs
Weekly tips

Click a card for details:

default · hermes

Chief of Staff

Strategy, ops cadence, calendar, email, 1:1s, faith, travel

jess

Jess

Research, thought leadership, deal hunter, breaking news

jackie

Jackie

Personal financial advisor (education, not fiduciary). GPT-5.5 via OpenAI Codex.

amit

Amit

College counselor for Gaby — pre-med, top-25, scholarships

What runs automatically

Filter by profile:

Single chatbot vs. this setup

One generic assistant

Context bleed between work, money, family
Repeats research every session
Hallucinated deal/news alerts
Manual morning routine every day
Cloud API costs scale with automations

George's Hermes setup

Isolated profiles with dedicated memory
RAG compounding + research logs
Python engines decide alerts; LLM discovers
16 cron jobs deliver briefs unattended
Hybrid models — local volume + Codex for finance quality

Integrations

Integration	Value
Google (gws)	Calendar, Gmail, Drive — agent reads/writes without browser
Granola	Past meeting notes for 1:1 prep — never confused with calendar
Telegram / Slack	Briefs where George already communicates
Qdrant RAG	Research from Tuesday informs Thursday's draft
Firecrawl + SearXNG	Structured scraping for deals and news

Key design decisions

Split agents by life domain

Work ops, research, finance, and family counseling need different safety rules and memory. One super-prompt cannot hold them all without bleed.

SOUL + AGENTS + USER + MEMORY files

Persona, playbook, facts, and decisions are version-controlled markdown — not buried in chat history. Agents read on demand.

Calendar (gws) ≠ Granola (notes)

Hard rule in every AGENTS.md: Google Calendar for what's next; Granola for what was said. Mixing them breaks briefings.

LLM researches; code decides alerts

Jess scripts find prices/headlines; Python engines (deals.py, news.py) decide delivery. Output [SILENT] when nothing qualifies.

Local-first for volume automations

CoS, Jess, and Amit cron jobs run on qwen3-coder:30b via Ollama — predictable cost for daily digests, deal scans, and triage.

Hybrid model routing — Jackie on Codex

Jackie uses gpt-5.5 via OpenAI Codex for interactive chat, auxiliary tasks, and both cron jobs (pinned). Finance analysis needs citation discipline local 30B can't match.

Dual delivery for Jackie cron

Portfolio report and news alerts deliver to Slack (Jackie channel) and Telegram (5946547528) so briefs aren't missed in one inbox.

Replicate this setup

Step 1 Install Hermes + Ollama — one profile first

Step 2 Write SOUL, AGENTS, USER — one page each; define tool routing

Step 3 Wire Google via gws — calendar + Gmail unlock most exec value

Step 4 One cron → morning brief → Telegram — prove unattended delivery

Step 5 Split profiles when domains fight for context

Step 6 Add deterministic scripts when LLM alerts lie or spam

Detailed architecture (interactive) → · Markdown