TechDesk AI — Autonomous Multi-Agent Social Intelligence System

TL;DR: An AI-native, event-driven system that monitors social platforms in real time, generates grounded responses using RAG, enforces safety via human-in-the-loop, and improves continuously using feedback loops (RLHF + bandits).

---

🖼️ Dashboard Preview

Real-time Human-in-the-Loop (HITL) dashboard for reviewing and approving AI-generated responses

---

🚀 Overview

TechDesk AI is a production-grade autonomous system designed to:

• Monitor social platforms (Reddit, LinkedIn, Twitter)
• Understand intent using LLMs
• Generate contextual, on-brand responses
• Route decisions through a multi-agent architecture
• Continuously improve using feedback loops

Unlike typical LLM apps, this system is built as a complete AI system, not just a prompt wrapper.

---

🧠 Core Idea

LLMs are probabilistic systems, not deterministic tools.

This system is designed to:

• Ground outputs → (RAG)
• Validate outputs → (HITL)
• Monitor behavior → (Audit logs)
• Improve over time → (RLHF + Bandits)

---

🏗️ Architecture — 7 Layers

┌─────────────────────────────────────────────┐
│ Perception                                 │
│ Reddit · LinkedIn · Twitter → Kafka        │
├─────────────────────────────────────────────┤
│ Understanding                              │
│ Intent · Sentiment · Entity extraction      │
├─────────────────────────────────────────────┤
│ Planning                                   │
│ LangGraph routing · Strategy selection      │
├─────────────────────────────────────────────┤
│ Memory                                     │
│ PostgreSQL + pgvector · Redis               │
├─────────────────────────────────────────────┤
│ Action                                     │
│ Response generation · Formatting            │
├─────────────────────────────────────────────┤
│ Safety                                     │
│ Filters · Toxicity detection · HITL         │
├─────────────────────────────────────────────┤
│ Observability                              │
│ Audit logs · RLHF · Strategy tracking       │
└─────────────────────────────────────────────┘

---

🤖 Multi-Agent System

            ┌───────────────┐
 Signal ──► │ Orchestrator  │
            └──────┬────────┘
                   │
      ┌────────────┼────────────┐
      ▼            ▼            ▼
 ┌──────────┐ ┌──────────┐ ┌──────────────┐
 │Engagement│ │  Crisis  │ │ContentCreator│
 │  Agent   │ │  Agent   │ │    Agent     │
 └──────────┘ └──────────┘ └──────────────┘
      │            │            │
      └────────────┴────────────┘
                   ▼
             ┌───────────┐
             │ Safety    │
             │  Gate     │
             └────┬──────┘
                  ▼
            ┌────────────┐
            │ HITL Queue │
            └────────────┘

Agents

• Orchestrator → intent classification + routing
• Engagement → normal responses (RAG + persona)
• Crisis → high-risk escalation
• ContentCreator → proactive/viral content

---

🔁 AI Engineering Workflow

1. Generate → LLM response
2. Ground → RAG retrieval
3. Validate → Human-in-the-loop
4. Log → Audit trail
5. Learn → RLHF + Bandits

---

⚙️ Production-Grade Optimizations

Prompt Engineering

• Structured prompts with output schemas
• Role-based constraints per agent

Hallucination Reduction

• RAG with pgvector
• Context injection before generation

Safety

• Keyword filtering + toxicity detection
• HITL approval for high-risk outputs

Observability

• Kafka event streaming
• Append-only audit logs for all LLM calls

Learning & Optimization

• RLHF preference collection
• Contextual bandit (epsilon-greedy)

---

🧰 Tech Stack

Layer	Technology
Backend	Python, FastAPI
Agents	LangGraph
LLM	Llama 3.3 70B (Groq API)
Streaming	Apache Kafka
Database	PostgreSQL + pgvector
Cache	Redis
Infra	Docker

---

📂 Project Structure

AI-Social-Agent/
├── services/
│   ├── perception/
│   ├── agents/
│   ├── safety/
│   ├── hitl/
│   ├── rag/
│   └── rlhf/
├── shared/
├── scripts/
├── infra/

---

🗄️ Database Schema

Table	Purpose
signals	Incoming social signals
actions	Agent outputs
knowledge_base	RAG data
audit_log	LLM call logs
preference_pairs	RLHF data

---

📡 Kafka Topics

Topic	Purpose
social.signals.raw	Raw signals
social.signals.classified	Classified signals
agent.actions.draft	Drafts
agent.actions.approved	Approved
agent.actions.published	Final outputs

---

🧪 Example Flow

1. Signal detected
2. Intent classified
3. Context retrieved
4. Agent selected
5. Response generated
6. Safety + HITL
7. Logged + feedback

---

🧠 Key Engineering Decisions

• LangGraph → explicit control
• Kafka → durable streaming
• pgvector → simple vector search
• Groq → fast inference
• fastembed → local embeddings

---

📊 What This Demonstrates

• AI-native system design
• Multi-agent orchestration
• Production-level safety
• Feedback-driven optimization

---

🔗 Repository

https://github.com/ankitnegi-dev/Techdesk-ai-social-agent

---

👤 Author

Ankit Negi

---

💬 Note to Reviewers

Happy to discuss:

• Architecture tradeoffs
• Prompt engineering strategy
• Failure cases & mitigations
• Scaling approach

"Build one layer at a time. Iterate on real data. The best agents are built by engineers who keep learning."