Product · v1.0

One system. Mines what to automate. Then builds it.

Routix is a vertically-integrated agentic platform. The observer sees the work, the synthesizer proposes the agent, the runtime executes it, the verifier guards it, and the analytics layer proves the ROI to your CFO — in one product, not five.

01 · Observer

A consented sensor for cross-app work.

Lightweight browser extension + desktop helper. Captures structured intents (click, navigate, paste, save) — never raw screenshots by default. PII is redacted on-device by Microsoft Presidio + a Routix-trained classifier before anything leaves the endpoint.

  • Open-source client; auditable in your VPC
  • Granular consent: per-app, per-class, per-user
  • Pause + replay-free mode for sensitive sessions
  • Air-gapped capture for regulated industries
02 · Pattern miner

Finds the work no one documented.

Sequence clustering, entity resolution across apps, and frequency-weighted ranking surface the top automations by hours-saved per week.

03 · Agent synthesizer

Generates the agent, not a flowchart.

Compiles observed patterns into LangGraph-style state machines with the exact tool calls, prompts, retry policies, and verifier checks pre-wired.

04 · Runtime

Durable execution on Temporal.

Every step is replayable. Retries are bounded. Costs are capped. Long-running workflows survive deploys.

05 · Verifier

An independent second opinion.

A separately trained model audits each consequential output for factuality, tone, and policy. Abstains when unsure.

06 · Reviewer UI

Humans in the loop, not in the way.

Triage queue with confidence sort, diff-style edits, and one-click approval. Decisions train the next-gen model.

[02] · CapabilitiesAll in V1.

Everything an enterprise buyer expects — already shipped.

Orchestration

Multi-agent

Planner → executor → verifier → writer, with shared memory and bounded retries.

Memory

Episodic + entity

Long-term per-user and per-org memory with Mem0/Letta-class retrieval.

Tool-use

MCP-native

200+ first-party connectors; register your own tools via SDK or MCP server.

Reasoning

Frontier models

Claude 4.x default, GPT-5 fallback, fine-tunes per workflow at scale.

Evals

Golden-set CI

Every PR runs regression on real customer scenarios. Drift blocks deploys.

Observability

Per-step traces

OpenTelemetry + Helicone; export to Datadog, Snowflake, BigQuery, S3.

Identity

SSO · SAML · SCIM

WorkOS-backed enterprise identity from day one. Per-action RBAC.

Policy

Budget + escalation

Spend caps, two-person rules, allow/blocklists, scoped tool authz.

Notifications

Slack · email · inbox

In-the-flow updates. Open a thread, approve, edit, escalate.

[03] · Inside the runtime

Watch a single task end-to-end.

routix · trace · pattern="renewal_brief" · run=#2841
# 00:00.000 — observer detected pattern (seen 12× in last 30 days)
pattern = "every Mon 9am, CS lead opens SFDC → filters renewals(30d) → manually drafts brief per acct"
candidates = 6 accounts in scope
# 00:00.041 — planner
plan:
  - query(salesforce, "renewals next 30 days")
  - enrich(clearbit, "accounts")
  - fetch(zendesk, "tickets last 90d")
  - draft(template="renewal_brief", voice="mara@helio")
  - verify(model="rx-verifier-v3", threshold=0.92)
  - notify(slack="#renewals", queue="cs-review")
# 00:00.812 — executor
ok  salesforce.query   → 6 rows · 240ms
ok  clearbit.enrich    → 6/6 enriched · 1.1s
ok  zendesk.fetch      → 41 tickets · 380ms
ok  openai.draft       → 6 drafts · 3.2s · $0.18
# 00:05.402 — verifier
pass 5/6   hold 1/6 (low factuality: 0.87)
# 00:05.611 — writer
post slack #renewals (5 briefs · attachments)
queue cs-review (1 brief · context attached)
# 00:05.804 — done · time saved this run: 47min · $4.20 spent

Built like infra. Sold like outcomes.