// projects.dir / multi-agent_llm_platform
A production multi-agent system built on Claude — specialized agents, persistent memory, real cloud infrastructure, and a Discord-native interface.
// what.it.is
This isn't an LLM chatbot. It's a coordinated team of AI agents — each with its own role, tools, and boundaries — operating against shared memory and a real production runtime. The platform was designed around a simple architectural belief: that coordination beats raw model power, and that the next generation of useful AI systems will look more like organizations than oracles.
The system runs continuously on dedicated cloud infrastructure. A central orchestrator agent ("Elena") receives requests, breaks them into delegable work, dispatches specialist agents in parallel, monitors progress, and reports back. Agents pass deliverables through a structured task log. Memory persists across sessions. Scheduled jobs fire automatically without human intervention.
The platform has been in continuous production since spring 2026.
// team.roster
Each agent has its own model selection, effort budget, turn limit, and explicit tool allowlist. Agents that don't need write access don't have it. Agents that don't need production credentials don't get them.
| Agent | Role |
|---|---|
| Elena | Discord-facing orchestrator and sole point of contact. Receives all requests, delegates to specialists, tracks progress, reports back. Reads seven identity files on every invocation. |
| Architect | System design. Consulted first on any new feature or initiative. |
| Backend Dev | Node.js, FastAPI, Express, database, auth. |
| Frontend Dev | React, Tailwind, TypeScript. |
| Python Dev | Owns the live Alpaca trading bot codebase. |
| QA Engineer | Required after every dev task. Sign-off gate. |
| Security Reviewer | Defensive code review. Read-only. Final gate before any feature is marked done. |
| Pentest Expert | Offensive web-app pentest, bug bounty methodology, report drafting. Companion to security reviewer. |
| Quant Advisor | Trading strategy consultant. Reviews performance and recommends parameter adjustments. |
// architecture.detail
Custom coordination layer. Agents run in parallel where work is independent, sequenced where dependencies exist. Architect designs; developers implement; QA tests; security reviews; orchestrator integrates and reports.
Anthropic Claude (sonnet and opus tiers) provides reasoning. Structured JSON outputs flow into downstream decisions: trading positions, news scoring, document drafting, compliance verification.
File-based memory system captures user preferences, project context, standing principles, and corrections. The platform learns across sessions, not within them. Typed and indexed for retrieval.
News API integration, scheduled job execution via APScheduler, structured persistence in SQLite. The system reads, decides, and acts without polling questions back to the human.
DigitalOcean VPS with systemd services, cron-driven heartbeats, log rotation, uptime monitoring. The platform doesn't stop when the developer steps away.
Environment-variable hygiene, signed API keys, no credentials in code or logs. Paper-trading vs production credentials are isolated.
Bidirectional Discord MCP integration. Requests come in via Discord; responses, attachments, scheduled reports, and morning heartbeats flow back the same way. No web dashboard required.
Structured logging across all components, daily summary aggregation, automated heartbeat reports to verify the system is alive every weekday morning.
// active.applications
The same platform runs multiple active workloads:
Paper-traded against the Alpaca API. The platform runs a nightly thesis-scoring pass over a 30-symbol watchlist, classifies each name as bullish, bearish, or neutral with a confidence score, then feeds those classifications into a screener and risk manager that places real orders. Trailing stops, position-size scaling by thesis confidence, and earnings-aware risk tightening all operate without human intervention.
The platform drafts production-quality HIPAA and other regulatory documents from primary source text. The workflow fetches the actual regulatory citations, builds a required-element mapping table, drafts prose against the mapping, and produces a self-verification report before any document reaches a human reviewer. Documents include Notice of Privacy Practices, Business Associate Agreements, patient intake forms, training acknowledgments, sanctions policies, and specialty add-ons across dental and mental health verticals, with primary care in active development.
Charged Earth Initiative architecture, capital-structure modeling, regulatory landscape research, and infrastructure planning all run through the same agent team.
// stack.detail
// public.scope
This is a private deployment. Live agent task logs, trading positions, business strategy artifacts, and client documents are not publicly accessible. What you see here is the architecture and the engineering, not the operating data.
For inquiries about applying this kind of platform to a specific decision domain — compliance monitoring, document review, customer-service triage, automated reporting — reach out via the contact form.