Roxy on Hermes Cloud
Bob’s separate private research assistant, hosted on Nous Hermes Cloud.
Roxy is Bob’s private research assistant on Nous Hermes Cloud. She collects Bob’s research, remembers how he works, and answers questions from his own source library.
Session results · August 30, 2026
Roxy carried the morning’s work list solo. Outcomes as Roxy reported them (library now 236 documents — 234 Substack + 2 events):
Partial / not done: Bob’s iCloud sync on the Mini is paused mid-diagnosis (handoff note in Dropbox — do not reset end-to-end encryption); the leftover unused local Hermes install was not checked; the basket scanners have not yet fired (built on a Sunday).
The foundation is live. Real Substack mail reaches Roxy, cited answers from the library are proven, and on-demand company-event ingestion is activated.
Bob’s separate private research assistant, hosted on Nous Hermes Cloud.
Bob can work with Roxy through Telegram and Hermes Desktop. Desktop uses Nous / Hermes Cloud as the backend — not a local agent. Daily setup is OK.
Honcho carries Bob’s preferences and working context. Woken and confirmed active (API + deriver) ahead of the August 14 session. Hosted tenants can sleep when idle; resume at the Honcho app if it naps again. Separate from the research-document library.
Roxy has a dedicated Gmail account. Bob’s Substack-only filter forwards publisher mail to Roxy. Real publisher mail is in the agent inbox, and library ingest from that mailbox is operational.
Bob owns a Supabase Pro database with hybrid search — word search plus search-by-meaning. As of August 30, the corpus holds about 234 Substack full-text documents plus two company-event records (236 total), with duplicate protection and daily automatic library ingest. 2,909 passages are embedded; hybrid search became the default on August 30.
Dropbox holds Roxy’s outputs, a practical rebuild kit, and the shared Inbox for large source files Bob wants Roxy to ingest. Inbox is a handoff, not a watched folder—Bob drops a file and asks.
Bob can hand Roxy a shareholder meeting, earnings call, investor day, conference, or management interview as captions, transcript, URL, or a Dropbox Inbox file. Official text wins; ElevenLabs Scribe v2 runs only when transcription is needed. Video is treated as audio—no slide or visual analysis. Two real events are already in the library. One artifact at a time.
Green components are live. Mail feeds the private library. Cited answers and honest misses are proven. Company events ingest on demand.
August 30 session (see results at the top of this page): smarter retrieval is finished — hybrid search is the live default with a calibrated honest-miss threshold. Both weekly digests verified delivered and useful. Two research basket scanners built; first fire Monday August 31. Cited answers from the live library remain proven (hit + honest miss + dedupe). Company event ingestion stays activated for on-demand use—captions first, Scribe only when needed, one artifact at a time. Library as of August 30: about 234 Substack documents and two company events (236 total). Still open: Bob’s iCloud sync on the Mini (paused mid-diagnosis) and confirming the leftover unused local Hermes install is deleted. Matt Pocock skills stay done at 7. Relationship mapping stays abandoned.
Bob closed this checklist on August 15. A question the library can answer returns a cited hit; a question outside the corpus returns an honest miss; re-ingest keeps a single document. Searchable library is still not complete research coverage.
7 safe Matt Pocock skills are live on production Roxy. This work is finished: no further installs are planned. The remaining skills from the set of 25 stay uninstalled; dangerous and caution-flagged skills stay blocked.
Roxy’s main model does the real thinking. Small background chores — reading a screenshot or image, summarizing long context, and naming sessions — do not need it. Those can run on a cheap, fast model instead, which lowers cost and latency without touching answer quality.
/review second-opinion reviewer. A separate agent reads Roxy’s recent work and criticizes it. Run it on the Codex subscription rail rather than a metered per-token charge.Pattern borrowed from Teknium’s own Hermes Agent setup (posted September 13, 2026): Google Gemini Flash assigned to vision, compression, and session titles, with a second model — Astra — giving a second perspective on /review. The three background chores use google/gemini-3.8-flash.
On the reviewer, one correction worth keeping. Teknium’s screenshot routes review through the Nous Portal, which is metered per token and is the most expensive line in that setup. Routing the same model through the Codex subscription instead is a flat fee already being paid — same reviewer, no per-token bill. Verify which rail a reviewer is actually on before calling it cheaper: a provider name can still land on a metered endpoint.
Planned, not done: these are configuration changes to Roxy’s auxiliary roles. None have been applied to Bob’s instance. Verify with a real screenshot read and a real conversation compression before calling it live.
These are possibilities, not commitments. The next product decision should follow Bob’s real use rather than a prebuilt phase tree.
A local Desktop would be a local agent with Bob’s disk permissions. James confirmed August 16 that the live Desktop uses Nous / Hermes Cloud as the backend. Daily setup is OK as-is. Do not rip Desktop out. Do not block the iMac move.
Leftover only: delete an unused local install if one is still on disk, so nobody can flip the app back to local. Residual is app chrome, not Mini disk. “Prompt for every local action” is still not a stock Hermes switch. Next session: look at Bob’s iCloud sync in the iMac → Mini move.
The corpus is no longer empty (about 76 Substack items plus two company events, dedupe working). Cited answers are proven. Trust still means Bob relying on those answers over time—not bulk storage alone.
Today Roxy finds documents by matching the words in Bob’s question against the words in the library. That works when the words line up, and fails silently when they don’t: ask about “datacenter power constraints” and a document that says “grid interconnect queues” never gets found—so Roxy honestly reports a miss on something the library actually contains.
The planned fix has three small parts, all inside Bob’s existing Supabase—no new services:
Shipped August 30 (Roxy’s report): hybrid search is the live default. The merge keeps word search primary and uses meaning search to backfill what words alone miss (a blended-ranking alternative was tested and rejected because it hurt exact-term lookups). The honest-miss threshold is calibrated at cosine 0.42 against the frozen gold set: vocabulary-mismatch 10/10, exact-term 5/5, out-of-corpus honest misses 4/4, author/date 3/3. 2,909 passages embedded. Live proof: a paraphrased question found the right passage at 0.538; a question the library doesn’t cover stayed an honest miss.
One concise, source-cited digest, emailed to Bob at 3:00 AM Monday, every line cited to a real document, honest about quiet weeks. Roxy’s first scheduled push. Built August 16: Bob opted in; first real send Monday Aug 17, 03:00 ET to rbcorman@gmail.com; a 65-line cited dry-run already went to James. Pause anytime with “pause digest.”
Analyze masked client portfolios for concentration, research exposure, and changes over time.
Fetch end-of-day prices on demand with a named source and as-of date, rather than building a standing market-data warehouse. Live August 16: on-demand EOD from Alpha Vantage (e.g. AAPL 2026-08-14 = $305.93). Raw tape only — split/dividend-adjusted close is premium-blocked, so the script refuses rather than guess.
Record dated theses with confidence, horizon, invalidation conditions, supporting sources, and later outcome review.
Test hypothetical positions separately from Bob’s real and client portfolios.
Optional teaching lab, not a production brain and not tomorrow’s work. Install LM Studio on Bob’s M4 Mini (64 GB) with Qwen 3.6 27B (MLX 8-bit). Qwen 3.8 27B is not released yet — do not download placeholder copies. LM Studio is the easy tool, not the fastest inference engine.
Start the local server only when Bob wants it; leave it off the rest of the time. This lab assumed Desktop-on-Mini talking to localhost. Cloud/Telegram Roxy cannot see that machine unless we later add a private tunnel — and a tunnel is a new hole. If the Desktop-on-Mini decision stays “Telegram only,” this lab stays off. If the local model is off, Roxy just says so and keeps using her Cloud model.