License: Source-available for portfolio review. Not licensed for commercial use or redistribution. See LICENSE.
Agentic deal intelligence for sales calls. Paste a transcript, get a structured deal brief in under a minute: stakeholders mapped, objections classified and matched against a sales playbook, compliance landmines flagged, and a follow-up email drafted — streamed to the screen as the pipeline runs.
🔗 Live demo: dealbrief-codejupiters-projects.vercel.app 🎥 30-second walkthrough: (Loom embed — coming soon)
LLM wrappers that take a transcript and call summarize() once are a dime a dozen. Real sales operations need something more structured: who's on the call, what they're objecting to, whether your AE just promised something legal is going to flag, and what to do tomorrow morning.
dealbrief is a five-step agentic pipeline that treats a sales call as a structured intelligence problem. Each step has a narrow job and produces typed output the next step can consume. The final synthesis sees everything and writes a brief a sales manager could hand a rep.
┌─────────────────────────────────────────────────────────────────────┐
│ TRANSCRIPT INPUT │
└──────────────────────────────────┬──────────────────────────────────┘
│
┌──────────────────────────┼──────────────────────────┐
▼ ▼ ▼
┌───────────┐ ┌───────────┐ ┌────────────┐
│ EXTRACT │ │ CLASSIFY │ │ COMPLIANCE │
│ │ │ │ │ │
│ Haiku │ │ Haiku │ │ Haiku │
│ │ │ │ │ │
│ People, │ │ Objection │ │ Unsupported│
│ companies,│ │ categories│ │ claims, │
│ products, │ │ + buying │ │ PII risks,│
│ amounts, │ │ signals │ │ regulatory │
│ timeline │ │ │ │ promises │
└─────┬─────┘ └─────┬─────┘ └─────┬──────┘
│ │ │
│ ▼ │
│ ┌───────────────┐ │
│ │ CROSS-REFERENCE│ │
│ │ │ │
│ │ Voyage AI │ │
│ │ embeddings │ │
│ │ │ │
│ │ Match objections│ │
│ │ against 12-entry│ │
│ │ sales playbook │ │
│ └───────┬────────┘ │
│ │ │
└──────────────────────────┼──────────────────────────┘
▼
┌────────────────┐
│ SYNTHESIZE │
│ │
│ Sonnet │
│ │
│ Streaming │
│ markdown brief │
│ + draft email │
└────────┬───────┘
│
▼
┌────────────────┐
│ DEAL BRIEF │
│ │
│ Stakeholders │
│ Objections │
│ Buying signals │
│ Compliance │
│ Next actions │
│ Follow-up email│
└────────────────┘
| Step | Model | Job | Output type |
|---|---|---|---|
| 1. Extract | Claude Haiku | Pull structured entities from raw transcript | People, companies, amounts, dates, products, decision criteria, timeline signals |
| 2. Classify | Claude Haiku | Categorize objections into a closed enum + surface buying signals | price | timing | technical | political | competitor | trust | other + severity + confidence |
| 3. Cross-reference | Voyage AI voyage-3-lite |
Match each objection summary against a hardcoded sales playbook via cosine similarity | Top-3 playbook patterns + recommended responses |
| 4. Compliance | Claude Haiku | Conservative scan for legal/trust landmines | Unsupported claims, certification misstatements, regulatory promises, PII commitments, competitor disparagement |
| 5. Synthesize | Claude Sonnet | Take everything above and write a senior-level deal brief | Streaming markdown — stakeholder map, top objections with responses, compliance flags, next actions, draft email |
Why three Haiku calls and one Sonnet? Haiku is fast and structured — generateObject with Zod schemas gives you typed, validated entities for cents per call. Sonnet only runs on the synthesis step where prose quality actually matters. The full pipeline costs roughly $0.02 per transcript.
Why embeddings instead of an LLM for cross-referencing? A playbook lookup is a similarity search, not a reasoning task. Calling an LLM to do "which of these 12 patterns matches" is overkill that adds 2–5 seconds of latency. Voyage AI's voyage-3-lite returns matches in ~500ms with a direct fetch to their API (the voyageai npm SDK has Turbopack bundling issues, so it's a thin custom client).
Why streaming? The synthesis step takes 30–70 seconds depending on transcript length. Streaming token-by-token over SSE means the brief starts appearing within ~2 seconds of synthesis starting. The UI feels alive instead of frozen.
Why a closed enum for objection categories? Open-ended classification produces inconsistent labels across runs. A closed set lets you build dashboards on top and compare deals at scale — which is what a real sales ops team would need.
Compliance bias toward flagging. Better to over-flag and have a rep ignore than miss an FDA misstatement that kills a deal during legal review. The compliance step has explicit prompts to err on the side of caution.
- Next.js 15 (App Router, Turbopack)
- TypeScript end-to-end with Zod for runtime schema validation
- Anthropic Claude — Haiku for extraction/classification/compliance, Sonnet for synthesis
- Voyage AI —
voyage-3-liteembeddings for playbook cross-referencing - Vercel AI SDK —
generateObjectfor structured output,streamTextfor synthesis - Tailwind CSS with custom HSL design tokens for the dark theme
- react-markdown for rendering the streaming brief
- Server-Sent Events for streaming the pipeline state and brief tokens to the client
git clone https://github.com/codejupiter/dealbrief.git
cd dealbrief
npm installCreate .env.local:
ANTHROPIC_API_KEY=sk-ant-...
VOYAGE_API_KEY=pa-...Then:
npm run devOpen localhost:3000, click Load sample, and pick one of the three included transcripts:
- Price objection — Mid-stage deal with a planted ROI claim that the compliance step should catch
- Easy close — Clean call with strong buying signals, no objections
- Compliance landmine — Healthtech sale where the rep makes an FDA misstatement, disparages a competitor without evidence, and promises PHI handling that conflicts with their actual product
The compliance landmine is the most interesting one — the synthesis step's draft email proactively corrects the FDA claim before the user's compliance team sees the original MSA.
dealbrief/
├── app/
│ ├── api/pipeline/route.ts # SSE endpoint, orchestrates all 5 steps
│ ├── page.tsx # Main UI
│ └── globals.css # Design tokens
├── components/
│ ├── Header.tsx
│ ├── TranscriptInput.tsx # Left pane — sample loader + textarea
│ ├── Pipeline.tsx # Right pane — live step status
│ └── Brief.tsx # Streaming markdown brief
├── lib/
│ ├── steps/
│ │ ├── extract.ts # Step 1
│ │ ├── classify.ts # Step 2
│ │ ├── crossref.ts # Step 3
│ │ ├── compliance.ts # Step 4
│ │ └── synthesize.ts # Step 5
│ ├── samples/ # Three sample transcripts
│ ├── playbook.ts # 12-entry sales playbook for cross-ref
│ ├── pipeline-types.ts # Shared types + step metadata
│ └── use-pipeline.ts # Client hook for SSE consumption
A short list of things that would move this from "demo" to "real product":
- Parallelize extract + compliance. They have no dependency on each other and currently run sequentially. Easy win, ~30% latency reduction.
- Prompt caching. The system prompts for each step are stable — caching them via Anthropic's prompt-caching API would cut input token costs by ~70% on repeat runs.
- LLM re-ranker on cross-reference. Vector similarity is fast but coarse. After embeddings narrow it to top-5 candidates, a small Haiku call to pick the best match would catch cases where the lexical similarity is high but the semantic fit isn't.
- Eval harness. A test suite with held-out transcripts and expected outputs (stakeholder lists, objection categories, compliance flag counts). Right now I tune thresholds by eyeballing three samples — that doesn't scale.
- CRM connectors. The natural next step. Pipe the structured output into Salesforce / HubSpot so reps don't have to copy-paste.
- Multi-call deal threading. A single call is one data point; deals span weeks. The bigger product is "show me how this deal evolved across the last four calls."
Built by Zoriah Cocio over a weekend as a portfolio piece. The goal was to demonstrate a real agentic pipeline — multiple LLM calls with structured handoffs, embeddings as a tool rather than a feature, and streaming UX — rather than yet another chatbot.
If you're hiring for full-stack or AI engineering roles and want to talk about how this was built: info@zoriahcocio.com.