Build Digital Partners
Standard frameworks crash or freeze during long-running tasks. Helm OS delivers a stateful, non-blocking runtime to deploy autonomous digital partners that maintain context and execute reliably for weeks.
Absolute Sovereignty Over Your Domain
Helm OS absorbs the brutal physics of distributed state, concurrency, and network resilience. You retain 100% control over the intelligence, methodology, and integrations. Build bespoke, multi-week digital partners without fighting the infrastructure.
Helm OS Kernel relies on Google Vertex but does not impose a vendor monopoly. Seamlessly plug your workflows into Google Vertex AI, OpenAI, Anthropic, or proprietary local open weights via standard API adapters without altering a single line of orchestration logic.
We provide the "Hollow Chassis", you provide the methodology. Domain developers are free to write custom workflows using standard, native asyncio Python: pure, flexible code governed by enterprise-grade execution boundaries.
Connect your cloud services, databases, proprietary microservices, and external SaaS APIs to your digital partner instantly.
Architecting Deep Continuity
Helm OS is an enterprise kernel that strictly decouples UI reflexes from deep reasoning enabling continuity for multi-week AI missions.
Bifurcated Persistence & Zero-Data-Loss
Helm OS eliminates decouples instant UI reflexes (System 1) from heavy background reasoning (System 2) via isolated Compute Sidecars. Additionally, its Bifurcated Persistence Engine (BPE) and Dual-Clock architecture automate secure, ultra-fast state saves in O(1) time (<20ms) without the latency or RCE vulnerabilities of standard massive JSON serialization.
# Helm OS: The Compute Sidecar Pattern
# The main orchestration loop never blocks. Heavy lifting is physically isolated.
from app.core.runtime import managed_workflow
@managed_workflow
async def autonomous_agent(context: WorkflowContext, file_bytes: bytes):
# 1. COMPUTE SIDECAR: Decouple System 1 (UI) from System 2 (Heavy Reasoning)
# Offloads intensive workloads to prevent event loop blocking and fatal crashes.
result = await context.compute.run_cpu_task(
heavy_data_parser,
file_bytes,
resources={"cpu": 1.0}
)
context.workflow_data.update(result)
# 2. BIFURCATED PERSISTENCE (BPE): Ultra-fast O(1) state saves (<20ms)
# Bypasses massive JSON serialization to eliminate latency and RCE vulnerabilities.
await context.save_checkpoint_delta()Micro-Step Idempotency, the in-memory storage fast-path
Designed for "Scale-to-Zero" economics, the system achieves zero-burden micro-step idempotency. The @managed_step decorator intercepts execution, hashes the inputs, and verifies the infrastructure checkpoint history. If preempted by a cloud provider, the system structurally enables resumption without duplicating expensive I/O operations.
# Helm OS: Micro-Step Idempotency
@managed_step(step_id="step_fetch_heavy_data")
async def _fetch_data(context: WorkflowContext, query: str) -> dict:
# If the Ray cluster crashes during a massive multi-week mission,
# this expensive I/O call will NEVER be executed twice.
return {"data": f"Results for {query}"}The Ghost Bridge
Standard AI agents crash when saving memory because live infrastructure, like network connections and background tasks, cannot be serialized into text. Helm OS solves this via the Ghost Bridge pattern, which isolates active system tools from raw data using hidden properties. When a developer runs await context.save(), the engine automatically strips these active connections to commit a clean snapshot. Upon waking or migrating servers, Helm OS restores the data and transparently re-binds it to fresh, functional infrastructure, ensuring seamless workflow continuity.
# Imports the authoritative state container that holds workflow_data and system services
from app.state.adapters import WorkflowContext
# Imports the kernel runtime decorator that registers the function with the orchestration engine
from app.core.runtime import managed_workflow
@managed_workflow
async def domain_specific_business_workflow(context: WorkflowContext, args: dict):
# 1. Domain developers use live, interactive system channels seamlessly
await context.notifications.send(["Initiating enterprise workflow..."])
# 2. Update standard database-bound state variables
context.workflow_data["current_milestone"] = "verification"
# 3. Save progress. The developer is completely isolated from the mechanics
# of state serialization, network filtering, or asset stripping.
await context.save()Decoupling enables safe, persistent enterprise AI
Helm OS digital partners decouple UI reflexes from heavy background reasoning. Powered by isolated Compute Sidecars and ultra-fast, secure state saves (<20ms), this architecture eliminates the latency and vulnerabilities of standard Python frameworks.
Defeats Python's GIL by offloading heavy reasoning to isolated Compute Sidecars in the Ray Cluster. Digital partners remain responsive to user interrupts even while executing complex, multi-week background missions.
Replaces risky, blocking JSON serialization with in-memory micro-step commits via a Dual-Clock architecture. If a cloud node preempts, Helm OS provides seamless resumption without duplicating processing.
Delivers a defensible Chain of Custody via the Semantic Flight Recorder. Operating as a transparent "Glass Box," it maps every LLM decision to exact trace IDs, while cryptographic provenance flags strictly prevent external lifecycle spoofing.
Helm OS eliminates agent orchestration boilerplate
Building enterprise grade agents requires massive boilerplate for database syncing, prompt engineering, and state routing. Helm OS eliminates this tax entirely by combining a "Schema-First," Zero-ETL data engine with Just-In-Time context orchestration. This allows developers to write clean, linear Python while the kernel automatically handles state persistence, production-ready GraphRAG, intent tracking, and task rehydration out of the box.
Semantic Born & Neurosymbolic Foundations
Simply define your business objects in a YAML file to auto-generate strict Python models. When your AI agent updates these objects in memory, a background process automatically translates them into graph relationships, generates vector embeddings, and commits them to Neo4j. This delivers production-ready GraphRAG and enforced data contracts entirely out-of-the-box—without requiring a single database query, migration script, or API call.
# ---------------------------------------------------------
# STEP 1: The Ontology (domain_schema.yaml)
# Define your entities and how they relate. Helm OS compiles
# this into strict Python models and Neo4j edge mappings.
# ---------------------------------------------------------
slots:
meta_client_id:
slot_uri: ergatis:retainedBy # Helm OS automatically translates this to a Graph Edge!
domain: Jobs
range: Clients
classes:
Jobs:
is_a: DomainEntity
attributes:
meta_job_title:
range: string
required: true
meta_client_id:
range: Clients
...
# ---------------------------------------------------------
# STEP 2: The Workflow (helpers_system_1.py)
# Update memory natively in Python. No Cypher queries,
# no API calls, no embedding scripts.
# ---------------------------------------------------------
async def create_new_job(context: WorkflowContext, job_id: str, client_id: str):
# 1. Define your standard Python dictionary
job_data = {
"id": job_id,
"canonical_uri": "ergatis:Jobs",
"meta_job_title": "Senior Cloud Architect",
"meta_client_id": client_id # The OS intercepts this Foreign Key
}
# 2. Package as a Semantic Entity
semantic_entity = {
"type": "SEMANTIC_ENTITY",
"entity": job_data
}
# 3. Commit to the O(1) Shadow Buffer
await context.memorize(semantic_entity)
# MAGIC: Once the buffer saturates, the 'Archivist' Sidecar takes over asynchronously.
# It generates a 768-D Vector Embedding for the job, strips 'meta_client_id',
# and projects a strict (Jobs)-[:retainedBy]->(Clients) edge into Neo4j!Augmented Context & The Single Voice
Deliver a seamless, single-threaded user experience with zero routing boilerplate. Helm OS maps UI signals to cognitive intent via YAML, runs background tasks silently, and injects this live telemetry into your digital partner's context Just-In-Time. The result is an 'Ask Me Anything' (AMA) Langgraph agent that manages suspended tasks—all without you writing a single custom prompt, state-recovery loop, or routing script.
# ---------------------------------------------------------
# STEP 1: Zero-Boilerplate Intent Mapping (domain_manifest.yaml)
# Map frontend UI clicks to Cognitive Primitives. No routing scripts required.
# ---------------------------------------------------------
manifest:
signal_mappings:
- raw_uri_pattern: "JOBS.SELECT"
cognitive_primitive: "FOCUS_SHIFT" # Helm OS automatically tracks what the user is looking at
target_entity_type: "Job"
meta:
description: "User is viewing a specific job requisition" # JIT injected into LLM context
# ---------------------------------------------------------
# STEP 2: Silent Background Execution (workflows.py)
# Background workers execute logic without colliding in the main chat UI.
# ---------------------------------------------------------
@global_registry.register_bg("analyze_candidates")
@managed_workflow
async def analyze_candidates_bg(context: WorkflowContext, args: dict):
# Perform heavy processing...
# Update the UI directly, but keep chat_state="off" to protect the AMA Agent's "Single Voice"
await context.notifications.send(
message_list=["System: Candidate pool optimized."],
chat_state="off",
done=True
)
# ---------------------------------------------------------
# STEP 3: The Omniscient AMA Agent (workflows.py)
# Zero state-recovery loops. Zero custom prompt injections.
# ---------------------------------------------------------
@global_registry.register_fg("ama_agent_langgraph", is_default=True)
@managed_workflow
async def ama_agent_runner(context: WorkflowContext, args: dict):
while True:
# 1. Wait for user input (Helm OS auto-suspends and recovers state if needed)
user_input = await context.notifications.wait_for_input()
# 2. Call the LLM.
# MAGIC: The InferenceGateway autonomously weaves the "FOCUS_SHIFT" from Step 1
# and the background states from Step 2 into the LLM's context window Just-In-Time.
success, response = await context.inference.invoke_chat(
context=context,
messages=[HumanMessage(content=user_input)]
)
# 3. Respond with the Single Voice
await context.notifications.send(
message_list=[response.content],
chat_state="on"
)Hybrid GraphRAG & Semantic Discovery
At the core of Helm OS is a persistent Long-Term Memory (LTM) powered by our proprietary, Zero-ETL Neo4j Hybrid GraphRAG engine. Rather than relying solely on basic vector similarity, Helm OS equips your digital partner with specialized Langgraph retrieval tools. This allows the agent to autonomously execute O(1) exact matches, perform multi-hop edge traversals, and run semantic vector searches to uncover hidden insights across complex datasets.
# ---------------------------------------------------------
# CAPABILITY 1: O(1) Exact Match Filtering
# The AMA agent uses this to instantly retrieve known entities without vector math.
# ---------------------------------------------------------
@tool
async def exact_match_entity_filter(entity_class_name: str, filter_key: str, filter_value: str, config: RunnableConfig):
# Executes a deterministic, O(1) database lookup for precise facts
results = await context.inference.retrieve_by_exact_match(
context=context,
entity_class_name=entity_class_name,
filter_key=filter_key,
filter_value=filter_value
)
return json.dumps(results)
# ---------------------------------------------------------
# CAPABILITY 2: Structural Graph Traversal
# The AMA agent uses this to "walk" graph edges and discover hidden relationships.
# ---------------------------------------------------------
@tool
async def traverse_graph_relationships(source_id: str, relationship_type: str, config: RunnableConfig):
# Bypasses vector similarity to execute a strict, multi-hop Cypher traversal
# (e.g., Finding all 'Candidates' connected to a specific 'Job' via the 'ranksCandidate' edge)
cypher_query = f"""
MATCH (src)-[r:{relationship_type}]-(target)
WHERE src.id = $source_id
RETURN properties(target)
"""
results = await context.inference.execute_cypher(
context=context,
query=cypher_query,
params={"source_id": source_id}
)
return json.dumps(results)
# ---------------------------------------------------------
# CAPABILITY 3: Hybrid Semantic Vector Search
# The AMA agent uses this for fuzzy, natural language discovery.
# ---------------------------------------------------------
@tool
async def search_specific_domain_entity(entity_class_name: str, query: str, config: RunnableConfig):
# Generates a 768-D embedding and executes a semantic similarity search
# across the Knowledge Graph, utilizing metadata filters to prevent dilution.
results = await context.inference.retrieve(
context=context,
query=query,
mode="HYBRID", # Combines Vector facts and Community insights
limit=20
)
return json.dumps(results)Engineered for Deep Cognitive Continuity
Helm OS eradicates instruction amnesia and chat pollution by structurally binding your agent to a persistent neurosymbolic graph and weaving real-time telemetry into a single, cohesive voice
Turn basic chatbots into true digital partners don't never forget context or rules. Your digital partner speaks with one all-knowing voice—delivering a seamless experience with zero routing or recovery code.
Give your digital partner a secure memory. Helm OS equips your Digital Partner to find facts, connect complex relationships, and run deep semantic searches.
Prevent your agent from losing focus during long, multi-week conversations. Helm OS automatically trims chat history and offloads heavy data files to protect the digital partner's attention span.
Accelerated builds. Enduring performance.
Helm OS automates complex infrastructure and memory management, empowering your developers to quickly build highly reliable Digital Partners that execute multi-week enterprise workflows without losing context.
Build with our Engineers
To help you transition to a true Cognitive OS, we're excited to offer a 6-week, fully sponsored engineering partnership for our upcoming v0.2.0 Beta. Participants will collaborate directly with our core engineers and receive exclusive, managed access to our underlying infrastructure—including a Shared Ray Cluster, Redis Event Bus, and partitioned Neo4j Graph—to seamlessly implement the Hollow Core architecture. To ensure we can provide the highest level of hands-on support and guidance, we are carefully selecting exactly three high-value enterprise use cases for this exclusive cohort.
STRICT ZERO-DATA RETENTION POLICY IN EFFECT. All inputs are cryptographically isolated, instantly routed, and strictly compliant with CCPA, GDPR, and SOC2 data sovereignty mandates. You are interacting with deterministic infrastructure, not a data-harvesting apparatus.