𝗜 𝗧𝗮𝘂𝗴𝗵𝘁 𝗛𝗶𝗻𝗱𝘀𝗶𝗴𝗵𝘁 𝘁𝗼 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿 𝗦𝗮𝗹𝗲𝘀 𝗖𝗮𝗹𝗹𝘀 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗡𝗼𝗶𝘀𝗲
Transcripts are too noisy to be memory.
If you feed an AI every word from a sales call, you get a junk drawer. You do not get intelligence. Most systems fail because they try to remember everything.
I built a Deal Intelligence Agent to solve this. It does not just transcribe. It remembers.
I used Next.js, FastAPI, Supabase, Hindsight, and Groq. Here is how I split the work:
- Supabase stores facts. It holds deals, meetings, and specific action items. Use this for exact queries.
- Hindsight stores memory. It holds the patterns and strategies. Use this for semantic recall.
I organized memory into three types to keep the agent useful:
• Episodic memory: What happened in a specific meeting. • Semantic memory: Patterns that emerge across multiple meetings. • Procedural memory: Strategies that actually work for a specific deal.
I also added a rule for timing. The agent must be humble.
If you have one meeting, the agent only records what happened. After two meetings, it looks for patterns. Only after three meetings does it suggest a strategy. This prevents the AI from making wrong assumptions too early.
The result is a massive difference in the pre-meeting brief.
A generic AI says: "Prepare for pricing objections."
My agent says: "Sarah Chen, the CFO, rejected the enterprise price in meeting two. She softened after you proposed phased pricing in meeting four. Lead with the phased structure today."
That is the difference between an assistant and a partner.
My key takeaways:
- Do not use transcripts as memory. Extract facts first.
- Use gates for memory types. Do not let the agent guess patterns after one call.
- Use two databases. Use relational stores for facts and vector stores for context.
- Names matter. Knowing a specific person's objection is better than knowing a "customer" objection.