Cambium + Pam: Infrastructure & Budget Proposal
Date: April 2, 2026 Author: Joshua Haynes
What This Covers
Infrastructure costs, time investment, and a budget for running Cambium and Pam — with room to scale from 3 deployments today to 15 over the next year.
What Is Cambium?
Cambium coordinates teams of AI agents. Think of it as an operating system for AI workers — not one chatbot answering questions, but multiple agents researching, analyzing, deciding, and acting together with safety checks throughout.
The design borrows from biology rather than corporate org charts:
- Signals — agents communicate through signals that strengthen or fade over time, like neurons firing. Stale information decays automatically.
- Immune system — detects abnormal agent behavior, tracks trust scores, and can quarantine a misbehaving agent without human intervention.
- Health monitoring — watches error rates, response times, and resource usage. Escalates when something goes wrong.
- Governance — some decisions need sign-off. Cambium enforces approval gates so high-stakes actions don’t happen without a human (or multiple agents) agreeing.
- Stigmergy — agents leave markers in a shared environment, like ants leaving pheromone trails. Other agents pick up on that work without needing direct communication.
- Lifecycle — agents are born from templates, do their job, and eventually shut down. When they do, their performance data gets “composted” back into the system so the next generation is better.
The whole thing is built in TypeScript, stores everything in PostgreSQL, and has a visual dashboard (Loom) for monitoring.
Why build this instead of using LangChain or CrewAI?
Those tools treat agents like employees in a hierarchy — rigid, scripted, fragile. That works for simple tasks. It falls apart when an agent fails mid-workflow, when you need 15 isolated deployments sharing one server, or when you want the system to actually learn from its mistakes. Cambium handles all of that.
What Is Pam?
Pam (Prediction Arbitrage Markets) does two things: market intelligence and cross-venue arbitrage. It runs as a deployment on Cambium.
Market Intelligence
Pam watches the news and maps events to prediction markets on Kalshi (CFTC-regulated U.S. exchange) and Polymarket (crypto-native). It scores how “eventful” a situation is — how much real evidence supports a given prediction — then evaluates trading opportunities across specialized strategies for politics, economics, legislation, and interest rates.
Every trade requires a human tapping “Approve” on Telegram. There is no autonomous trading.
Cross-Venue Arbitrage
When the same event is priced differently on Kalshi vs. Polymarket, that’s free money (minus fees and execution risk). Pam scans both venues, matches equivalent markets using semantic analysis, computes the spread, sizes positions using Kelly Criterion (capped at 8% of bankroll), and executes both sides atomically within 250–500ms.
It also catches intra-venue inconsistencies — like when YES + NO contracts on the same event don’t add up to 100%.
How Pam Uses Cambium
Different agents handle different jobs: signal discovery across both venues, strategy evaluation, market matching for arbitrage, human approval routing, and a research agent (“Ralph”) that runs backtests to optimize parameters in a sandbox. Cambium’s health layer watches API connectivity to both exchanges; the immune system flags anomalous market data; governance enforces the human-in-the-loop rule on every trade.
Safety
- Defaults to dry-run mode. Live trading requires explicit
--live --confirm. - Every trade needs manual approval on Telegram. No exceptions.
- Research runs against historical data only.
- Arbitrage is all-or-nothing — one failed leg cancels the whole trade.
Current Deployments
| Deployment | What It Does | Status |
|---|---|---|
| Obadiah | AI chief of staff — daily briefs, task coordination. Eight specialist personas covering health, finance, innovation, business strategy, security, QA, marketing, and UX. | Active development |
| Pam | Prediction market intelligence + cross-venue arbitrage (Kalshi and Polymarket) | Active |
| Community Health Agent | Discord community governance — conflict prediction, health scoring, adaptive moderation, onboarding | Alpha |
The Community Health Agent
This is Cambium’s first commercial product. Standard Discord bots delete messages with banned words. The Community Health Agent is an intelligent governance layer:
- Spots brewing conflicts before they blow up, using signal intensity and decay patterns
- Generates a real-time health score for the community based on sentiment, engagement, and moderation load
- Learns what “healthy” means for each specific community instead of applying generic rules
- Runs personalized onboarding for new members
- Enforces configurable governance workflows (role assignments, channel access, escalations)
Right now it’s running in report-only mode on Joshua’s Discord — observing and scoring, not acting. That’s deliberate: validate the models first, automate later.
Infrastructure
Three phases. The application code stays the same at each step — only the AI model endpoint changes.
Phase 1: Managed API (Now)
Run Cambium locally on a mid-range Linux box. Use cloud-hosted AI models.
Server: AMD Ryzen 7 7700X, 64 GB DDR5, 2 TB NVMe, Ubuntu. ~$1,100.
AI models: Together.ai and Groq serving Gemma 4 27B, with Claude as emergency fallback. Cambium’s provider system handles failover automatically.
| Monthly Cost | |
|---|---|
| AI model APIs | $75 – $150 |
| Electricity, DNS, services | $25 – $30 |
| Total | $100 – $180 |
Phase 2: Self-Hosted Cloud GPU (Month 3–6)
Move inference to a RunPod RTX 4090 you control, running Ollama. Costs slightly more (~$150–175/mo) but gives you model version control and no rate limits.
Phase 3: Local GPU (Month 6–12)
Buy an RTX 4090 ($1,800), add RAM and a bigger power supply (~$300 more). Monthly costs drop to $70–130 — mostly electricity and a small Claude API bill for complex reasoning.
Cumulative hardware after Phase 3: ~$3,200.
If 15 concurrent deployments ever bottleneck on inference, a second GPU and CPU upgrade runs ~$6,000. Not needed initially.
Time Investment
How This Gets Built
One person (Joshua) working with AI coding assistants (Claude Code, Codex). The AI handles bulk code generation; Joshua handles architecture, design decisions, strategy, and review. Output rate is closer to a small team’s, but the cost is one person’s time.
Time Already Invested
Development started August 2025. With AI doing the heavy lifting on code generation, the actual hours at the keyboard are lower than the codebase size suggests:
| System | Hours | Notes |
|---|---|---|
| Cambium (18 packages, ~26K LOC) | ~200 – 280 | Architecture, directing AI, reviewing output, testing |
| Pam (~8K LOC) | ~80 – 120 | Strategy design, parameter tuning, backtest validation |
| Obadiah (8 personas) | ~30 – 50 | Persona design, workflow architecture |
| Community Health Agent (alpha) | ~20 – 30 | Discord integration, health scoring prototype |
| Total | ~330 – 480 hrs | Since August 2025 |
Ongoing Commitment
| Activity | Hours/Week |
|---|---|
| Platform maintenance | 3 – 4 |
| Pam monitoring & trade approvals | 2 – 3 |
| Community Health Agent (alpha to beta) | 6 – 8 |
| Loom dashboard | 2 – 3 |
| New deployments | 1 – 2 |
| Total | 14 – 20 |
What That Time Is Worth
Berlin market rates as opportunity cost:
| Hourly | Monthly (17 hrs/wk) | Annual | |
|---|---|---|---|
| Senior engineer | €85 – €120 | €5,780 – €8,160 | €69,360 – €97,920 |
| AI/ML specialist | €120 – €180 | €8,160 – €12,240 | €97,920 – €146,880 |
| Blended (80/20) | €100 – €135 | €6,800 – €9,180 | €81,600 – €110,160 |
The AI assistant costs show up in the API line items. There’s no separate AI budget.
Total Budget
Team: 1 person + AI. No salaries, contractors, or overhead.
Year 1
| Low | High | |
|---|---|---|
| Hardware (server + GPU) | $3,000 | $3,300 |
| Operating (12 months) | $1,200 | $2,100 |
| Hard costs | $4,200 | $5,400 |
| Joshua’s time (opportunity cost) | €81,600 | €110,160 |
| Total with time | ~€85,800 | ~€115,560 |
Year 2 (steady state, local inference)
| Low | High | |
|---|---|---|
| Hardware | $0 | $500 |
| Operating (12 months) | $840 | $2,060 |
| Hard costs | $840 | $2,560 |
| Joshua’s time | €81,600 | €110,160 |
| Total with time | ~€82,440 | ~€112,720 |
Break-Even
Infrastructure breaks even at 25 communities on the $49/mo Growth tier. Time investment breaks even around 150–200 paying communities. At 1,000 communities, projected ARR is ~$445K at 85%+ gross margins.
Revenue
Community Health Agent (SaaS)
| Phase | When | What | Revenue |
|---|---|---|---|
| 1 | Month 1–3 | Alpha to beta, first 10 customers | ~$1K MRR |
| 2 | Month 3–6 | Add impact reporting for DAOs | ~$5K MRR |
| 3 | Month 6–12 | DAO governance tooling | ~$15–25K MRR |
| 4 | Month 12–24 | Mental health community platform | ~$50K+ MRR |
Pricing: Free / $49 / $149 / $499 / Enterprise (custom).
Pam (Trading Returns)
Variable. Depends on market conditions and arbitrage opportunities between Kalshi and Polymarket. Structured as asymmetric risk: conservative position sizing caps downside, spread opportunities provide upside. Not modeled as predictable revenue.
Risks
| Risk | Mitigation |
|---|---|
| LLM costs spike | Local inference by Phase 3 removes the dependency |
| Hardware failure | PostgreSQL backups + Cambium’s checkpoint/recovery system |
| Gemma 4 not good enough for complex work | Claude API fallback built in from day one |
| Time commitment too high | Revenue from Community Health Agent funds the work; phased approach limits exposure |
| Prediction market regulation changes | Human-in-the-loop only; Kalshi is CFTC-regulated |
| Competition | Composting data flywheel + 26K LOC of hardened orchestration code create real switching costs |
Technical Overview
Stack
| Languages | TypeScript (strict), Python |
| Runtime | Node.js 22+ |
| Database | PostgreSQL 16, Drizzle ORM |
| Monorepo | pnpm + Turborepo |
| Tests | Vitest, pytest |
| UI | Next.js 15, React 19, React Flow, Tailwind 4 |
| AI models | Gemma 4 (local/cloud), Claude (fallback) |
| Model server | Ollama |
Codebase
| Cambium | 18 packages, ~26,000 lines TypeScript |
| Pam | ~8,000 lines Python |
| Tests | 25 passing, 7 integration (need live LLM) |
| Specs | 16 PRD documents |
Local AI (Gemma 4)
Gemma 4 27B fits on a single RTX 4090 at 4-bit quantization (18 GB VRAM). Open weights, no licensing fees, commercial use allowed. Cambium already has a built-in Ollama adapter. A smaller 4B variant handles routine tasks cheaply. For the 20% of work that needs top-tier reasoning, Claude API fills the gap.