Scattered operational truth
Scouting lists, investor notes, portfolio updates, and menu intelligence live across spreadsheets, CRM dumps, and enrichment tools that never agree.
Axion
Workflow Engine
A governed, multi-tenant platform for designing, testing, publishing, and operating graph-based AI workflows — where AI assists and humans decide.
The problem
Point tools cannot be the operating system
Accelerators and adjacent B2B teams need orchestration and institutional memory — not another chatbot, and not another copy of Apollo or PitchBook.
Inconsistent evaluation
Thesis fit, propensity scores, and report drafts depend on who ran the last model — with little evidence, lineage, or replayability.
Ungoverned AI risk
Direct provider SDKs and ad-hoc agents promote unverified outputs into systems of record, with no review gate, cost ceiling, or audit trail.
The answer
One engine. Four business motions.
Shared runtime, connector framework, AI gateway, evidence model, lineage, review system, and audit layer. Templates become governed apps on one stack.
Startup
Source and score companies against a living thesis — discover, enrich, dedupe, then promote only what reviewers confirm.
Fundraising
Build investor shortlists with exclusions, match reasons, and campaign-ready handoff segments — not cold-send spam.
Portfolio
Turn evidence into investor-ready reports with founder review and admin confirmation before publish.
FoodScout
Close the gap between menus, ERP catalogue, and purchase history → ranked cross-sell opportunities.
How a workflow runs
From draft graph to governed write-back
Workflows are data, not code. The graph runtime is a durable control plane — not fire-and-forget scripts.
Design the graph
Compose reusable nodes as versioned workflow data — not ad-hoc scripts.
Validate and publish
Sandbox first, then publish an immutable workflow version with a graph hash.
Execute with workers
Durable queue, retry, and pause — every step carries correlation and telemetry.
Bind evidence and lineage
Scores require citations. Field lineage keeps every decision explainable.
Human verification gate
Approve, reject, or correct before promotion.
Write canonical records
Confirmed outputs land in Axis. Downstream systems get one-way final copies.
AI you can trust
AI is never the final authority
Schema validation, evidence policy, and human review decide what becomes canonical. Every meaningful decision should be explainable.
AI Gateway only
All model access is centralised — prompt versions, structured JSON, repair, evidence policy, cost, and redacted traces. No direct provider SDKs from domain code.
Evidence-bound scoring
Fit scores and synthesis require citations. Confidence labels from VERY_LOW to VERY_HIGH travel with every meaningful output.
Governed hybrid agents
Supervisors may plan, act, observe, and replan with allow-listed tools. Promotion, export, and source-of-truth writes stay outside the model’s authority.
Institutional learning
Corrections become labelled feedback linked to workflow, prompt, and output versions — so the platform improves without silent drift.
Connect safely
Integrations as governed plugins
External systems stay connectors or downstream copies — never competing masters of Axis operational truth.
Capability-first adapters
Workflows depend on stable capabilities like b2b.company_discovery and web.crawl — not vendor payload shapes.
Apollo & web intelligence
Business-only enrichment, search, and crawl/parse pipelines with governed credentials via the secrets broker.
Files, email, ERP
Ingest artefacts and operational systems without turning Axis into a clone of transactional systems of record.
HubSpot one-way sync
Downstream final-copy only. Axis owns workflows, runs, prompts, evidence, reviews, and derived opportunities.
Built for operators
Safe enough for production ops
Fail-closed posture in staging and production: live providers, vaulted secrets, and agents where required. Mocks stay in local, test, or explicit sandbox simulation.
Multi-tenant isolation
Tenant scope on every run, review, connector credential, and audit event.
RBAC & review roles
Builders, operators, reviewers, founders, investor viewers, and admins — each with a clear lane.
Cost & quota preflight
Caps and ceilings before live spend. Sandbox simulation keeps experiments off production budgets.
Retry, DLQ, audit
Durable step state, dead-letter recovery, correlation IDs, and a compliance-ready audit log.
Every side effect must be safe to retryEvery tenant must be isolatedEvery review-worthy action must be auditable
Next step
Open the operating surface
Sign in to run workflows, review gates, and connector health.