Suki Pulse · Camp Run Hermes Agent Hackathon
Everything we built for CAMP / RUN: Build Overnight with Hermes Agent (DEVCON Manila × Avtica × Amihan, Avtica office, Makati, October 2, 2026).
Suki Pulse turns Suki Mart's customer complaints into operational fixes, in one click inside Hermes Desktop. Hermes runs the loop: TypeSafe Jev labels every complaint in about half a second, our MCP tools trace each cause into operations (riders, stock, suppliers, the loyalty ledger), and Hermes fixes both sides after asking first. Track: Open Innovation (Business Operations + Customer Experience).
Live links
One site, one Vercel project (suki-pulse). The old jev-live-race.vercel.app and jev-decision-race.vercel.app links (and their QR codes in the deck) were retired on October 4, 2026.
What's in this folder
| Folder |
What it is |
Made by |
Camp-Run-with-Hermes-Agent/ |
The three layers: mcp-server/server.py (10 suki tools), skills/suki-pulse/, desktop-plugin/suki-pulse/, on the Suki Mart sandbox (data/store.db, 221,142 rows). A git submodule: our fork of the organizers' starter kit. |
Hermes |
suki-pulse-site/ |
The one deployed site (suki-pulse.vercel.app): / the dashboard of the full Jev run (Hermes), /race/ the live race page + api/ functions with guardrails (passcode, same-origin, per-visitor limits, 600-char text, batches of 20), /benchmark/ the recorded race |
Hermes, Claude Code |
jev-speed/ |
bench.py (Jev vs Claude Haiku 4.5 benchmark), data.json (measured results), build.py + template.html (recorded page), app.py + live.html (local live race on 127.0.0.1:8790), build_live_deploy.py |
Claude Code |
videos/ |
Demo films: remotion/ (final cut out/suki-pulse-draft-a-sound.mp4), film/, shots/ |
Claude Code (video session) |
pitch/ |
script/ (pitch scripts and notes), deck/ (source files of the Claude Slides deck), qr/ (QR codes + make_qr.py, each verified to decode) |
Claude Code |
docs/ |
PLAN.md (idea, QR decodes, data findings, build plan), our Hermes setup guide (Hermes Setup Guide.pdf), the prep checklist page, Suki Pulse write-ups |
Claude Code, Hermes |
tools/ |
probe_suki.py (planted problems), pitch_facts.py (every pitch number), reset_store.py (reset the data while Hermes holds it open), verify_pulse.py, promo_check.py, decode_qr.py |
Claude Code |
assets/ |
Photos of the event slides (challenge, QR codes, judging) |
us |
Run it
# Hermes: register the MCP server (absolute path; restart Hermes Desktop after)
hermes mcp add suki --command uv --args run C:\dev\devcon\hermes-hackathon\Camp-Run-with-Hermes-Agent\mcp-server\server.py
# Reset the Suki Mart data (works even while Hermes has it open)
python tools\reset_store.py
# Local live race (Jev vs Haiku), then open http://127.0.0.1:8790
cd jev-speed; python app.py
# Re-measure and rebuild the recorded page
cd jev-speed; python bench.py; python bench.py --haiku-burst; python build.py
# Rebuild the live race page into the site, then commit and push: the suki-pulse
# Vercel project builds suki-pulse-site/ from main (Root Directory suki-pulse-site)
cd jev-speed; python build_live_deploy.py
Secrets
No key is stored in this folder. The TypeSafe key lives in C:\dev\.secrets\devcon-typesafe-jev.key, the OpenCode Zen key in Hermes's own .env (%LOCALAPPDATA%\hermes\.env), and the deployed live race read TYPESAFE_API_KEY, OPENCODE_ZEN_API_KEY and DEMO_PASSCODE from encrypted environment variables on the suki-pulse Vercel project. All three were removed on October 3, 2026, so the paid routes now refuse every request; set them again and redeploy to bring the race back. data/store.db is Syncthing-ignored because the MCP server writes to it live.
Measured numbers (October 2, 2026)
- Jev median 586 ms vs Claude Haiku 4.5 2.29 s on the same 40 complaints and 8 questions (3.9x faster); $0.000023 vs $0.000829 per decision (35x cheaper).
- All 694 open complaints across 12 branches: 64.7 s, 388,465 input tokens, $0.0163, zero fallbacks.
- Suki Mart is fictional sandbox data from the organizers' starter kit.