OpsDesk — Operational Intelligence Layer for Service Businesses
A B2B SaaS platform that builds a structured operational intelligence layer on top of email and calendar — starting with law firm intake revenue protection, expanding toward a governed company brain.
More screenshots
OpsDesk
The core insight: service businesses don't fail from lack of data. They fail because work doesn't reliably happen when it should. Prospects aren't followed up. Consults go without a next step. Billable actions get missed. The raw material is all there — in inboxes, calendar events, threads — but the meaning is scattered. OpsDesk reconstructs it.
What I built
The entry wedge is a law firm intake intelligence layer. The system connects to email and calendar (read-only), reconstructs each prospect's timeline from first inquiry through consult to engagement, and detects exactly where the workflow broke down.
The detection model is built on an 8-state intake state machine: New inquiry → Responded → Consult scheduled → Consult completed → Awaiting follow-up → Engagement in progress → Stalled → Closed. Every risk rule maps to a state transition that should have happened but didn't. When intake stalls, the system surfaces the at-risk matter, shows what triggered the flag, and drafts the next action for one-click approval.
Beyond the intake monitor, the platform handles the full pre-signing workflow stack — post-consult follow-up, engagement letter chase, no-show recovery, referral source loop closure, and the compliance-grade conflict check detection that carries legal risk (not just revenue risk) when missed.
The detection approach is deliberately layered: pure pattern matching where the signal is unambiguous (conflict check language, DocuSign links), LLM inference where context is required (fee discussions, referral signals, prospect intent). Claude handles the inference layer with explicit prompts — the system explains every flag, never produces unattributable alerts.
Architecture
The 6-phase product roadmap is built on a 7-layer platform architecture — not a feature roadmap:
| Layer | Function |
|---|---|
| Entity memory | Prospect, owner, matter, communication thread |
| Event timeline | Inbound emails, outbound replies, calendar events, state changes |
| Risk detection | Pattern-based and LLM-based detection on state transitions |
| Playbooks | Encoded expected behavior: triggers → steps → timing → escalation |
| Action layer | Draft emails, reminders, routing — human approves before any send |
| Role-aware views | Partner / associate / intake manager / admin — same data, different contexts |
| Cross-workflow coordination | Full company brain: reasoning across intake, billing, matter management simultaneously |
The governing rule: every feature request must map to one of these 7 layers. If it doesn't strengthen a layer, it doesn't ship — even if a customer asks for it. This is how a narrow intake tool compounds into a platform without becoming a custom service business.
Technical decisions
Inbox-native, not CRM-native. The CRM is always incomplete — it reflects declared intent, not ground truth. The inbox shows what actually happened. OpsDesk reads the inbox.
Manual-first for Phase 1. The first 3–5 firms are deeply monitored implementations, not automated deployments. Every false positive and missed detection is product feedback. The manual phase converts real intake behavior — firm-specific timing patterns, vocabulary, edge cases — into the rules and thresholds that later automate correctly. It's the fastest path to accuracy, not a compromise.
Human-in-loop by architecture, not just policy. No auto-send, ever. Users approve every action. The product earns trust incrementally — detection first, drafting second, execution third. The trust ramp is the onboarding flow.
Build platform beneath the wedge. The internal entity/event/state/risk/playbook model uses no law-firm-specific vocabulary. The law firm UI is a thin vertical layer over a horizontal architecture. Phase 3 expansion (accounting, financial advisory, healthcare) requires workflow templates, not re-architecture.
Stack
Claude · OpenAI · RAG over email/calendar · Postgres · Make.com · Python · FastAPI · React · Phase 3+: Temporal · Neo4j · vector DB · agent orchestration