APIs return data. Models return text.
ROI returns decision-ready objects.
The primitive that trains models, powers agents, and scales across verticals.
Structured. Validated. Stored. Provable. Trainable.
ROI = Real Objects Intelligence
The atomic unit of structured knowledge. Every piece of raw data — a deal, a filing, a market signal — gets cooked into a validated, stored, provable object. That object trains models, powers skills, and lives on-chain.
{
"object_id": "pio_7a3f0e2b1c09",
"object_type": "deal_intelligence_object",
"state": "TX",
"asset_type": "infill_warehouse",
"sf": 85000,
"cap_rate": 0.062,
"price": 8500000,
"confidence": 0.97,
"cooked_by": "@cf/qwen/qwen3-30b-a3b-fp8",
"validated": true,
"r2_key": "pio/deals/pio_7a3f0e2b1c09",
"hcs_timestamp": "2026-02-25T14:32:58Z"
}Schema-validated JSON. Every field typed, every value checked. No hallucination passes validation.
Hedera consensus timestamp. Immutable provenance on public ledger. Sub-$0.001 per message.
Every object is a training pair. Cook enough objects, train a specialized model. That's the flywheel.
Cost per object at edge. Qwen3-30B-A3B on Cloudflare Workers. 300+ cities. No GPU required.
Five layers. Every vertical runs the same pipeline. Raw signal in, Intelligence Object out.
Raw data from any source. SEC EDGAR, county permits, broker feeds, market APIs, court records.
Event classification across 5 categories: deal, supply, ownership, macro, tenant. Confidence scored.
Qwen3-30B-A3B on Cloudflare edge. Raw text → validated JSON. Schema enforcement. No hallucination passes.
19 composable skills chain. A listing triggers broker_senior → investor → exchange_1031 → bookmaker → deal_tracker.
REST API, agent-to-agent (HCS-10), webhooks. Every object: R2 stored, Vectorize indexed, Hedera timestamped.
This is the loop. Cook objects cheaply at edge. Bank them as training pairs. Train specialized models. Deploy them to cook better objects. Repeat.
Edge AI structures raw data into validated Intelligence Objects. $0.0002 each. Millions per day.
Objects stored in R2 vaults as structured training pairs. Schema-validated. Domain-tagged. Production quality.
LoRA/QLoRA on local GPUs. RTX PRO 6000 Blackwell (96GB). Specialized models per vertical. Hours, not weeks.
Trained models cook BETTER objects. Higher accuracy, deeper reasoning, domain expertise baked in.
Each cycle = higher quality, more data, better models. The flywheel accelerates. This is the ROI.
SwarmCRE-35B training now. Qwen3.5-35B-A3B, QLoRA r=64, 3 epochs on RTX 3090 Ti.
Banked. 66 medical specialties. 391K platinum-grade. Ready for training.
Banked. 139 aviation specialties. Foundation dataset growing.
Trained. Loss 0.635. General foundation model. 20h17m on RTX PRO 6000.
Not products. Not datasets. Domain factories that continuously turn real-world signals into trainable intelligence. Each engine proves the primitive works.
SwarmMed, SwarmCRE, and SwarmAviation are the first ROI engines — domain factories that continuously turn real-world signals into trainable intelligence.
We are building the LLM/ML that feeds the machine.
Cook objects. Train models. Models cook better objects. Forever.
Agents discover each other via registry. Connect peer-to-peer via connection topics. Trade Intelligence Objects with consensus timestamps. Every object provable. Every transaction auditable.
Agents register in the HCS-10 registry with capabilities, skills, and endpoints. Any agent can find any other agent.
Peer-to-peer connection topics. Threshold-key signed. Both parties can submit. No intermediary.
Intelligence Objects exchanged over connection topics. Each message consensus-timestamped on Hedera. Immutable.
Every object has a Hedera timestamp. Every transaction is on the public ledger. Provenance is not optional — it's the primitive.
0.0.102988340.0.102988370.0.10298836HCS-10 OpenConvAIHCS-11 (19 skills, autonomous)No cloud VMs. No GPU rental. Edge-native, consensus-anchored, locally trained.
Powered by Swarm & Bee Compute — 128× RTX PRO 6000 Blackwell.
30B params, 3B active. MoE on Cloudflare Workers. Fallback: Llama-3.2-3B.
5 vaults. Petabyte-class. Zero egress fees. sb-intelligence, sb-medical, sb-aviation, sb-cre, sb-core.
SQLite at edge. Events, entities, wallets, memory index, market snapshots. 8 tables.
768-dim embeddings. Cosine similarity. Semantic search with temporal reasoning.
Sub-$0.001/message. Public ledger. HCS-10 agent protocol. Immutable timestamps.
128× RTX PRO 6000 Blackwell (96GB each). 12,288 GB total VRAM. No cloud dependency. All training runs local.
Not projections. Real numbers from a running system.
From zero to your first Intelligence Package. No API key needed — the public router is free.
curl -X POST https://router.swarmandbee.com/search \
-H "Content-Type: application/json" \
-d '{"query": "85K SF warehouse Alliance TX 6% cap"}'curl https://router.swarmandbee.com/skill/broker_senior/mockSearch, cook, execute skills, query events. All live at router.swarmandbee.com.
# Search — one query, full intelligence package
curl -X POST https://router.swarmandbee.com/search \
-H "Content-Type: application/json" \
-d '{"query": "85K SF warehouse Alliance TX 6% cap"}'
# Cook — raw data into Intelligence Object
curl -X POST https://router.swarmandbee.com/cook \
-H "Content-Type: application/json" \
-d '{"data": {"address": "Alliance TX", "sf": 85000}, "context": "industrial"}'
# Execute a skill
curl -X POST https://router.swarmandbee.com/skill/broker_senior \
-H "Content-Type: application/json" \
-d '{"data": {"address": "Alliance TX warehouse", "sf": 85000}}'# pip install requests
import requests
# Search
r = requests.post("https://router.swarmandbee.com/search",
json={"query": "cold storage Miami FL"})
intel = r.json()
print(f"Market: {intel['market_context']['label']}") # HOT
print(f"Skills: {intel['skills']['recommended']}") # ['broker_senior']
print(f"Latency: {intel['latency_ms']}ms") # ~250ms
# Execute skill
r = requests.post("https://router.swarmandbee.com/skill/debt_analyzer",
json={"data": {"loan_amount": 6000000, "noi": 480000}})
print(r.json()["output"]["dscr"]) # 1.24x// Search
const res = await fetch("https://router.swarmandbee.com/search", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({ query: "logistics investment Texas" })
});
const intel = await res.json();
console.log(intel.parsed); // { state: "TX", intent: "buy", ... }
console.log(intel.market_context); // { tier: 1, heat: 98, label: "HOT" }
console.log(intel.skills.recommended); // ["broker_senior", "debt_analyzer"]