Periskope AI Agent

1. Overview

The Periskope AI Agent is an AI-powered automation layer embedded into the Periskope WhatsApp-based customer management platform. Its primary purpose is to assist or fully handle inbound customer conversations, reduce manual agent workload, and ensure consistent, policy-aligned responses at scale.

The agent operates in real-time WhatsApp conversations, with deep integration into Periskope’s internal systems such as:

• Ticketing
• Private internal notes
• Customer context
• Business knowledge bases
• Message history and metadata

Unlike a generic chatbot, the AI Agent is context-aware, stateful, and action-capable, meaning it can:

• Understand ongoing conversation history
• Retrieve relevant organizational knowledge
• Decide when to act vs escalate
• Perform internal actions (ticket creation, notes, tagging) via tools

2. Customer Problem Being Solved

Before the AI Agent:

• Human agents had to manually respond to repetitive questions
• Context switching between chats caused delays and errors
• Internal notes and tickets were inconsistently created
• Knowledge reuse across conversations was limited
• Scaling support teams directly increased cost

The AI Agent addresses these by:

• Handling repetitive and first-level queries automatically
• Maintaining consistent answers using shared organizational knowledge
• Creating internal artifacts (tickets, notes) without human intervention
• Escalating only when confidence is low or rules require it
• Reducing response time and operational load

3. Core Capabilities

3.1 Context-Aware Conversation Handling

• Processes full WhatsApp chat history
• Understands message order, sender roles, and prior resolutions
• Maintains short-term conversational memory per chat

3.2 Knowledge-Augmented Responses (RAG)

• Retrieves relevant documents, FAQs, and self-learned contexts
• Uses similarity search instead of keyword matching
• Grounds responses strictly in retrieved context when required

3.3 Tool-Based Actions

The agent can invoke internal tools, including:

• Ticket creation
• Private internal notes for human agents
• Context tagging and metadata updates

Each tool has:

• Explicit schema
• Strict usage rules
• Validation at execution time

Hall Of Fame

4. Evolution & Progression

Phase 1: Static AI Replies and Basic RAG

• Initial version was a simple LLM prompt + response
• No actions, no memory, no tools
• Limited usefulness beyond basic replies
• I implemented a Retrieval-Augmented Generation (RAG) system:
  • Business documents and FAQs are vectorized
  • Chat context is embedded and queried
  • Relevant context is fetched (cosine + semantic similarity) and then injected into the prompt

Limitations identified:

• Hallucinations
• No alignment with internal workflows
• No escalation mechanism

Phase 2: Tool-Enabled Agent

I introduced function calling / tool invocation, allowing the AI to:

• Create tickets
• Add private notes
• Follow predefined operational rules

Key improvements built:

• Explicit tool schemas
• Rule-based instructions for when tools can be used
• Separation between customer-visible responses and internal actions

This transformed the AI from a chatbot into an operational agent.

Phase 3: Improved RAG & Self-Learning

In this iteration I massively improved the context retrieval by:

• Parsing more conversation context by condensing it for the AI
• Using a combination of vector search, cosine similarity and falling back to semantic search for retrieval.

I added self-learned context generation, where:

• A training AI listens to messages being unflagged by internal team members and schedules a training job after 5 minutes.
• In the training job, it fetches the last 10 messages in the conversation and finds out how the internal member solved the customer query
• It then generates high-quality Q&A pairs or modifies existing ones
• These are stored and reused in future chats

Challenges identified:

• Self-learned contexts could be incorrect
• Overconfidence in low-quality learned data

Phase 4: Confidence, Re-ranking & Guardrails

To address reliability issues, I worked on:

a) Context Re-ranking

• Multiple retrieved contexts are scored
• Only the most relevant are injected
• Reduces noise and hallucination risk

b) Confidence Scoring

• Agent responses are evaluated post-generation
• Confidence thresholds decide:
  • Respond automatically
  • Ask a clarification
  • Escalate to a human

c) Human Review Loop

• Self-learned contexts can be reviewed before activation
• Prevents polluted knowledge bases
• Adds accountability and trust

Phase 5: Evaluation & Scoring Framework

I built an AI evaluation pipeline:

• Responses are judged against system rules
• Scored across multiple dimensions (granular scoring out of 100)
• Helps compare prompt versions and agent behaviors

This enabled:

• Regression detection
• Safer prompt changes
• Objective quality tracking over time

5. Current Architecture (High-Level)

Input

• WhatsApp message
• Chat history
• Customer metadata

Processing

1. Context retrieval (RAG)
2. Prompt assembly with rules and constraints
3. Tool availability injection
4. Model generation
5. Tool execution (if applicable)
6. Post-response evaluation

Output

• Customer-facing message
• Optional internal actions
• Confidence score
• Logs for audit and training

6. What I Built & Owned

I was directly responsible for:

• Designing the AI Agent architecture
• Implementing tool-based actions with strict schemas
• Building the RAG pipeline and self-learning system
• Identifying and fixing hallucination and overconfidence issues
• Designing re-ranking and confidence-scoring strategies
• Creating the evaluation and scoring framework
• Iteratively refining prompts based on real customer feedback
• Bridging AI behavior with real operational workflows

I shaped how the AI reasons, acts, and fails safely.

7. Current State

As of now, the Periskope AI Agent:

• Handles real customer conversations reliably
• Acts within well-defined operational boundaries
• Learns cautiously with human oversight
• Is measurable, debuggable, and improvable
• Reduces support workload while maintaining quality

It has moved from a simple LLM feature to a production-grade AI system embedded deeply into Periskope’s core product.

8. Impact

• Faster response times for customers
• Reduced manual effort for support teams
• Consistent answers across conversations
• Lower onboarding cost for new agents
• A scalable foundation for future AI-driven workflows