Projects⁽²⁷⁾

A listener chooses a piece and a language, then follows a story-first path that sets the scene and introduces a few concrete things to hear before or during the music. Context and guided listening arrive in small steps, so newcomers build confidence without being dropped into theory-heavy analysis.

A single web app fronts the whole clinic. Check-in writes a patient onto a shared queue that every screen subscribes to over server-sent events, so the front desk, providers, and billing all see the same state in real time: no polling, no stale views. Charting, telehealth, and claims hang off that same record.

Firstmate is the sole captain-facing surface: the captain gives it a goal, and it routes the work through persistent secondmates that each hold a scoped responsibility over time. Those secondmates coordinate projects and outcomes while subordinate workers stay internal, so the public model stays legible without exposing every execution identity. This work builds on Kun Chen's open-source Firstmate project: https://github.com/kunchenguid/firstmate.

A kiosk or phone flow captures demographics, insurance, and pre-visit answers, runs OCR + eligibility on the insurance card, and drops a structured, validated record straight onto the live queue: pushing a real-time arrival alert to staff. The provider sees complete data before the patient is roomed.

Work is spread across dedicated machines instead of one box. Each runner is isolated: its own sessions, its own browser, its own context: so dozens of agents work in parallel across many repos without contention or context bleed, and nothing pollutes the human's workstation.

A pull request triggers a review on a fresh, throwaway container (so secrets stay local). After posting findings, CODY captures what it missed, researches a detector improvement, and opens a PR against itself: so review precision compounds with every pass instead of plateauing.

A question fans out to three retrievers at once: LLM reasoning, a graph database for relationships, and vector search for meaning: then a synthesizer merges the hits into one cited answer and flags any coverage gaps. No single method has to be right alone.

Each source gets a distinct voice so you can tell notifications apart by ear. New voices are made by blending existing ones in embedding (x-vector) space: averaging the speaker vectors, not splicing audio: which is what keeps any single donor's identity out of the result.

A launcher scores every available account by remaining headroom and picks the best one per call, failing over automatically when one is exhausted: so a single rate limit never silences the system. Accounts are discovered from disk, so adding one is dropping a file, not editing code.

A scheduler pulls clinical encounters and runs them through rule packs plus an LLM to flag likely coding mistakes: mis-leveled visits, wrong patient class, missing codes. Each finding carries a confidence band and a suggested fix, queued for a human biller rather than auto-applied.

Inbound patient texts hit an AI layer that answers routine questions (hours, services, wait times) and runs post-visit follow-up. Anything clinical is detected and escalated to a human immediately: the safety guardrail is the point, so it never gives medical advice.

A sensor reads every agent's vitals each tick: context pressure, rate-limit status, liveness: and scores risk on a smooth curve (no threshold cliffs). When risk crosses a line it offers a treatment (compaction, transfer, resurrection) that the agent can refuse; consent is never overridden.

It aggregates many streams: commitments, deals, infra health, spend: and runs material-change detection so it only speaks when something actually moved. The output is one narrative brief, not a wall of metrics, so it never decays into wallpaper.

A fleet is governed like a state: an executive decides and acts, a legislature ratifies rules as versioned law, and a judiciary adjudicates. Term limits force every artifact to survive rotation, and a 'librarian' garbage-collects stale law so the rulebook stays small as the system grows.

When a session is born it durably records its identity and where it lives. After a crash, reboot, or terminal-server failure, the engine replays that record to restore the exact conversations and working context: deterministically, with no manual re-selection.

A card photo is preprocessed and read by a document-AI model, with a fallback OCR engine if confidence is low. It extracts payer, member ID, and plan, then feeds a real-time eligibility check: all on infrastructure the clinic controls, so card images never leave their walls.

Live terminal sessions from many machines stream into one configurable grid over an encrypted channel: CCTV for an agent fleet. It's read-only by default; taking control of any single pane is an explicit, gated action, so watching can never accidentally become writing.

It discovers acquisition targets, researches each into a sourced profile, picks the strongest personalization angle, and drafts outreach: which then passes through an adversarial loop: a writer drafts, a critic attacks, and a simulated buyer reacts, stripping generic and creepy language before anything sends.