AI Agent OrchestrationDeployed in Production

AI-Canvas

Social Media Automation SaaS Platform via Agentic Orchestration

Live Deployment GitHub Repo

The Problem

Creating high-converting LinkedIn content at scale is extremely manual. Generic LLM wrappers fail because they lack brand voice adaptation, multi-step critique, and intent-aware reasoning.

Why it matters

B2B founders spend 10+ hours a week drafting contextually relevant social content. A deterministic, automated pipeline drastically cuts CAC.

Who is affected

Founders, DevRel Engineers, and SaaS Marketing Teams.

Architecture & System Design

Frontend

Next.js App Router for highly responsive interactive UI with Socket.IO client for live token streaming.

Backend Processing

Node.js environment running LangGraph state machines for deterministic agent routing.

ML Pipeline

LangSmith handles LLM observability and trace abstraction; multiple LLMs deployed specifically via prompt routing.

Architectural Reasoning

Chose LangGraph over CrewAI for highly deterministic logic loops (Critique -> Rewrite) which guarantees content isn't published until it passes a strict internal heuristic.

Alternatives Considered

Standard direct-API calls were considered but quickly led to hallucinated or 'robotic' output. Multi-agent debate was strictly required.

ML & Technical Deep Dive

Model Selection & Training

Core Architecture

GPT-4o (Reasoning & Orchestration) + Claude 3.5 Sonnet (Creative Execution)

Training Methodology

Zero-shot with heavily engineered system prompts + dynamic few-shot retrieval via RAG context injection.

Dataset

Fine-tuning context windows built upon 500+ top-performing proprietary LinkedIn posts (A/B tested data).

Preprocessing Pipeline

.01Semantic chunking of previous posts
.02Vector storage via Pinecone for historical tone retrieval
.03RAG query expansion

Evaluation Metrics

94%

Generation Coherence

<150ms TTFT

Token Streaming Latency

85%

Time Saved/Post

Technical Challenges

Problem: Agentic Loop Hallucination

Solution: Implemented strict LangSmith tracing to detect infinite loops and added max-iteration thresholds in LangGraph.

Problem: High Latency in Multi-step Orchestration

Solution: Pipelined asynchronous LLM calls and pushed token-by-token generation straight to the Next.js frontend via WebSockets.

Core Features

Intent-Aware Generation

Uses user goals, target audience, and context to generate highly specific content.

Real-time Streaming Logs

Socket.IO feeds live agent logic status to the user (e.g. 'Agent analyzing...', 'Critiquing draft...')

Human-in-the-loop (HITL)

Requires manual approval before scheduling, keeping final editorial power with the user.

Results & Impact

Users transitioned from spending 2 hours brainstorming/drafting one post to generating 5 targeted posts in 10 minutes.

Currently actively used by beta users to maintain professional brand presence effortlessly.

80% Content Creation Time Reduction

Sub-150ms TTFT (Time To First Token)

Takeaways & Learnings

What I Learned

Mastered LLM Observability; LangSmith is mandatory for production agentic systems to trace exactly why an agent chose a specific tool.

Trade-Offs Made

Sacrificed minor end-to-end latency to force the LLM to undergo a 'Critique & Refine' sub-graph, increasing quality exponentially.

Future Improvements

Migrate the long-term semantic memory from Pinecone to pgvector for infrastructure consolidation.

Tech Stack Foundation

Frontend

Next.js
TailwindCSS
Framer Motion

Backend

Node.js
Socket.IO
Supabase

ML / AI

LangChain
LangGraph
LangSmith
OpenAI API

Tools

Pinecone
Vercel

Interested in this architecture?

Let's talk about how I can build something similar for your team.

All Projects Get in Touch