DealPulse

An autonomous enterprise deal-risk agent with a reusable MCP skill fabric: surfaces hidden blockers, dispatches modular tools, and hands off with SHA-256-verified context.

Comment

Inspiration

Enterprise B2B sales teams lose billions of dollars every quarter not because they lose deals to competitors, but because hidden operational blockers surface too late. We've seen it firsthand: a $350K deal collapses 48 hours before close because legal suddenly demands a non-standard BaFin EU data residency clause something that had never been surfaced or escalated.

The root cause is always the same: context loss. When an Account Executive escalates a deal blocker to Legal, Security, or the Deal Desk, critical context fragments across informal Slack messages, unrecorded calls, and undocumented decisions. There is no structured incident transfer protocol. There is no cryptographic audit trail. There is no autonomous agent capable of seeing the full picture across CRM data, call transcripts, compliance risk, and economic buyer engagement simultaneously.

We built DealPulse because we believe the first enterprise AI agent to solve deal context loss: the one that can autonomously parse a 15-turn sales call, extract 8 MEDDPICC qualification gaps, dispatch the right specialist skills, and hand off with a verifiable incident packet which will unlock an entirely new category of revenue intelligence.

What it does

DealPulse is an autonomous enterprise deal-risk orchestration platform. It ingests multi-speaker sales call transcripts and CRM pipeline context to:

Extract 8-Dimension MEDDPICC Scores: Continuously evaluates Metrics, Economic Buyer, Decision Criteria, Decision Process, Paper Process, Implicate Pain, Champion, and Competition from 12–15 turn sales conversations.

Dispatch 9 Modular MCP Skills: A decoupled skill registry conforming to the Model Context Protocol standard dispatches 9 production-ready tools:

crm.get_opportunity: Ingests CRM opportunity topology, stakeholder maps, ARR, and pipeline stage transcript.analyze: Parses multi-speaker call turns, identifies speakers, and calculates sentiment friction meddpicc.extract: Extracts 8 MEDDPICC methodology dimensions with confidence scores and quote citations crm.update_stage: Synchronizes stage transitions (e.g. "At Risk / Legal Review") and sets P1 risk flags in Salesforce crm.add_next_steps: Appends milestone action items, assigned owners, and cutoff deadlines to CRM tasks email.draft_ae_brief: Composes executive outreach emails and 1-page ROI justification models ticket.create_legal_review: Dispatches expedited compliance review tickets (#LEG-9941) with clause redlines and SLA targets handoff.route_specialist: Cryptographically seals incident transfer packets and routes execution to specialist agents handoff.to_human: Escalates deal blockers to human Deal Desk authority (VP RevOps) and freezes autonomous actions Generate Verifiable Handoff Packets (Handoffs OS): Compiles tamper-evident incident transfer notes with SHA-256 checksums, verbatim transcript citations, open MEDDPICC gaps, and prior decisions — guaranteeing zero context loss when escalating to Legal, Security, or human executives.

Maintain an Immutable Action Ledger: Every skill execution records correlation IDs, millisecond latencies, input payloads, and output schemas for a strict, auditable trail.

Optionally Enrich with Gemini Flash : 100% offline and deterministic by default. With a VITE_GEMINI_API_KEY, the orchestrator uses Gemini Flash for dynamic, context-aware chat phrasing while keeping deterministic heuristics as the source of truth.

Two demo scenarios are included:

Scenario A ($185K ARR) — Unengaged Economic Buyer risk with mid-range deal score (54/100) Scenario B ($350K ARR, Critical) — Late-stage legal block triggering full 9-skill MCP dispatch, legal ticket creation, CRM milestone sync, and human escalation with SHA-256 handoff (84/100)

How we built it

Architecture: Supervisor + Decoupled Skill Fabric

The system is built around a strict separation of concerns between the orchestration supervisor and the skill execution layer.

src/engine/orchestrator.ts : The DealPulseOrchestrator is the autonomous supervisor. It receives a deal scenario and deterministically executes a multi-step diagnostic pipeline: fetching CRM opportunity topology, analyzing transcript sentiment and stress, extracting MEDDPICC dimensions with confidence intervals, dispatching legal review tickets for Paper Process blockers, drafting executive ROI emails for unengaged economic buyers, synchronizing CRM tasks, and routing to specialists or human escalation as appropriate. It is a pure state machine — every transition is logged.

src/mcp/registry.ts : The MCP Skill Registry houses 9 modular tools each with a formal MCP-compliant schema (name, namespace, description, inputSchema, outputSchema, category). Skills are namespaced by domain: crm, audio_nlp, sales_intel, comms, compliance, orchestration. Any external MCP client (Claude Desktop, Cursor, Copilot Studio) could consume these schemas directly.

src/engine/heuristicAnalyzer.ts : Deterministic NLP heuristics parse multi-speaker transcript turns for keyword signals (security posture, legal resistance, economic buyer presence, competition mentions) with no runtime LLM dependency. This ensures fast, reliable, reproducible results.

src/engine/geminiService.ts : A conditional wrapper that enriches orchestrator chat messages with Gemini Flash natural language when an API key is present. Falls back to deterministic template responses. The heuristics remain the source of truth regardless.

src/components/ : React 19 component layer with a 3-column layout: Chat Console (left), MEDDPICC Risk Panel (center), and a tabbed Platform Panel (right) with MCP Skills, Handoffs OS, and Action Ledger views.

Stack: TypeScript + React 19 + Vite + Tailwind CSS + Cloudflare Pages. Zero backend. No database. Cold-start in under 2 seconds. Deployed via Cloudflare Pages with git integration.

Challenges we ran into

Designing an MCP-first skill registry without a running MCP server. The hackathon constraint was a browser-side SPA with no backend. We solved this by implementing a faithful MCP schema specification layer in TypeScript, each skill has a fully-typed inputSchema and outputSchema matching MCP JSON-RPC conventions, so the registry is externalization-ready for Stage 2 even though it currently runs in-process.

Making the handoff verifiably tamper-evident without a cryptographic backend. We implemented SHA-256 checksums using the Web Crypto API (browser-native, no Node.js dependency) over a canonical JSON serialization of the handoff payload. This produces a verifiable fingerprint that changes if any field of the incident transfer note is modified.

Keeping the orchestration deterministic while enabling optional Gemini enrichment. We needed the demo to work reliably in offline/cold environments while allowing optional Gemini Flash integration. The solution was a hard separation: heuristic analysis always runs first and sets all risk scores and action decisions; Gemini is only called afterward to rephrase the chat message text. If the Gemini call fails or the key is missing, the deterministic text is used verbatim.

Balancing UI information density with clarity. The 3-panel layout needed to surface MEDDPICC scores, MCP skill schemas, live ledger entries, and handoff packets simultaneously without overwhelming the user. We iterated on tab groupings and collapsible detail modals to create a "progressive disclosure" pattern.

Accomplishments that we're proud of

A fully functional autonomous orchestration pipeline that executes a 9-step deal diagnostic across two realistic enterprise scenarios without any manual intervention. Verifiable SHA-256 handoff packets — a novel mechanism for agent-to-agent-to-human context transfer that eliminates information loss at escalation boundaries. A genuinely reusable MCP skill registry with 9 production-quality tool schemas that could be plugged directly into any MCP-compatible client. Instant Human Escalation & Auto-Actions Freeze — single-click escalation to human Deal Desk authority with automatic ledger logging and automated execution locking. Zero backend, zero database, zero cold-start — the entire system loads and executes in under 2 seconds from a global CDN edge. Optional Gemini enrichment with deterministic fallback — the system is reliable and demo-safe regardless of API key availability.

What we learned

The Model Context Protocol is a genuinely powerful standard for skill composability. Designing each tool with a formal schema first (before writing any implementation logic) forced us to think clearly about input/output contracts and made the registry immediately understandable to any agent runtime. Cryptographic handoffs are underused in agentic systems. SHA-256 checksums on incident transfer notes are a simple, powerful primitive for establishing trust and auditability at handoff boundaries — and they add negligible overhead. Deterministic heuristics and LLMs are complementary, not competing. Using heuristics for structured extraction and LLMs only for natural language formatting proved to be the right division of labor for reliability + expressiveness. Enterprise sales context is deeply domain-rich. MEDDPICC is just one framework. Real Stage-2 work would integrate Challenger Sale, SPIN, and custom qualification scorecards per vertical.

What's next for DealPulse

Milestone Capability
Stage 2.1 Freshworks MCP & Agent Studio — Native integration with Freshsales CRM and Freshservice deal desk workflows
Stage 2.2 ElevenLabs AE Voice Briefs — 60-second audio pre-call briefings highlighting active MEDDPICC gaps
Stage 2.3 Bi-Directional MCP Server Endpoints — Expose DealPulse skills over JSON-RPC HTTP/SSE for external agent clients
Stage 2.4 Live CRM & Conversational Connectors — Two-way sync with Salesforce, HubSpot, and Gong/Chorus webhooks
Stage 2.5 Automated Ground Truth Evals — Benchmark suite testing MEDDPICC extraction precision across 500+ enterprise deals

Built With

Share this project:

Updates