Inspiration
What it does
The idea for Agentic AI was born from observing how real-world sales and customer success teams often lose track of context across multiple touchpoints — calls, texts, emails, and CRM notes. A single prospect might talk to different agents over days or weeks, yet the conversation starts from zero every time.
Inspired by conversational assistants like OpenAI’s ChatGPT memory and human-like sales AI tools such as Drift and Gong, I wanted to build a system that remembers every interaction and suggests the next best action (NBA) automatically.
🧩 What I Built
Agentic AI is a multi-channel conversational intelligence backend that uses AI memory and reasoning to understand the context, detect intent shifts, and recommend follow-up actions.
The system:
Stores and retrieves memory from Supermemory API.
Tracks all customer touchpoints — calls, emails, SMS, CRM notes.
Detects objections like “too expensive” or “just browsing”.
Generates human-quality follow-ups (SMS, Email, or Call script).
Escalates to a human agent when needed with a summary + recommended script.
Example workflow:
A 7-minute call is transcribed.
Agentic AI detects an objection.
It sends 2 SMS + 1 Email and books a 2nd call, ensuring continuity.
🧠 What I Learned
This project deepened my understanding of:
Context retention and memory embeddings for long-term AI interactions.
How to build multi-modal reasoning systems that combine NLP, structured data, and historical logs.
Practical AI deployment with FastAPI, Supermemory, and LLM-driven intent detection.
Managing stateful, evolving dialogue using architecture similar to cognitive agents.
I also explored how Quality of Inference (QoI) applies to decision accuracy in multi-turn conversational flows.
🏗️ How I Built It ⚙️ Architecture Overview graph TD A[FastAPI Backend] --> B[Supermemory Vector Store] B --> C[Intent Detector (LLM-based)] C --> D[Next Best Action Engine] D --> E[Message Generator (SMS, Email, Script)] E --> F[CRM Integration Layer]
🧩 Core Components
Backend Framework: FastAPI + Uvicorn
Memory Management: Supermemory.ai
Language Model Logic: Custom intent classification (call → objection → escalation)
Data Simulation: generate_calls.py creates realistic multi-turn demo transcripts
Storage: Local + Supermemory persistent memory
Analytics: Detects keyword-based intent drift using scoring heuristics
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Agentic_AI
🚧 Challenges Faced
Memory Alignment: Early Supermemory API versions had inconsistencies in how filters were passed — requiring deep inspection of the client SDK.
AI Generation Control: Balancing between creativity and factual precision in LLM outputs (e.g., avoiding generic “Dear [Prospect’s Name]” messages).
Multi-turn Data Simulation: Generating realistic synthetic calls that reflect real intent shifts across 10+ touchpoints.
Context Drift: Ensuring older interactions are not forgotten when prioritizing the newest message.
🧭 Next Steps
Integrate real audio transcription (Whisper / AssemblyAI) for live call ingestion.
Add LLM-driven style mimicry for each prospect (writing tone, word choice).
Expand NBA logic to include auto-calendaring (Calendly API) and CRM sync (HubSpot/Zendesk).
Build an interactive frontend dashboard to visualize the entire conversation graph.
✨ Takeaway
Agentic AI isn’t just an assistant — it’s a memory-powered sales brain. It learns, adapts, and reasons about human intent, allowing agents to act faster, stay consistent, and build trust with every follow-up.
Log in or sign up for Devpost to join the conversation.