MicroBrand Voice Engine: An Agentic Copywriter with Persistent Memory, Guardrails & Observability

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

💡 What I Built

Solo founders, artisan makers, and indie side-hustlers waste hours adapting product copy across multiple channels:

• A mindful, storytelling Shopify product description.
• A punchy Instagram Reel / TikTok script paired with spoken audio.
• A feature-bulleted, conversion-oriented Amazon listing.

Standard one-shot LLM prompts fail because chatbots forget brand voice guidelines between sessions and default to generic hype phrases ("cheap", "insane deal", "limited time") that ruin the brand identity of bespoke, minimalist, or craft brands.

MicroBrand Voice Engine is an agentic marketing co-pilot built to solve this:

1. Persistent Brand Memory: Uses assistant memory to remember brand tone, signature slogans, and negative constraints permanently.
2. Deterministic Agent Workflow: A multi-stage LangGraph pipeline that retrieves memory, generates channel-specific copy via open-weight inference, and passes output through a compliance guardrail node.
3. Audio Previews: Synthesizes instant spoken voiceovers for social video scripts using ElevenLabs.
4. Production Observability: Full execution tracing via Sentry Agent Tracing to monitor token budgets, model latency, and runtime errors.

💸 Who Would Pay For It?

• Boutique Shopify / Etsy Sellers: Solo creators who cannot afford a dedicated copywriter or agency ($1,000+/month) but need consistent, multi-channel product marketing.
• Freelance Social Media Managers: Marketers managing multiple brand identities who need isolated memory workspaces and compliance guardrails that prevent brand voice drift.

---

📹 Demo

• Live Web Application: https://microbrand-engine.onrender.com/ (Hosted on Render Free Tier with 24/7 automated ping keep-alive; loads instantly)

💻 Code

The complete source code, agent graphs, and deployment blueprints are open-source on GitHub:

Anubrata-Mallick / microbrand-engine

Open-source agentic copywriting engine with persistent memory and voice synthesis for modern e-commerce brands.

An autonomous marketing engine that transforms raw product specs into brand-compliant, multi-platform copy and audio voiceovers with zero voice drift.

---

⚡ At a Glance

The Dilemma Boutique makers and solo founders waste hours manually writing tailored copy for Shopify, Instagram, and Amazon. Standard LLM prompts forget brand personas between sessions and constantly invent generic buzzwords ("bargain", "crazy sale") that degrade luxury or artisan brands.

The Solution MicroBrand Voice Engine stores brand tone, signature slogans, and negative constraints in persistent assistant memory, runs a 3-stage critique state machine, generates targeted copy, and creates realistic spoken voiceovers for social reels.

Live Streamlit UI: Persistent Brand Workspace • Multi-Channel Copy • Spoken Voiceover Playback

---

---

🏆 Hackathon & Sponsor Shoutouts

This project was built for the Hacktoberfest 2026 Open-Source AI Challenge (Week 2: Side Hustle). Huge shoutout to the event partners whose developer platforms made building and running this engine…

View on GitHub

---

🛠️ How I Built It

The engine is engineered around open-source AI frameworks, open-weight model architectures, and modular cloud primitives:

1. Agent Orchestration (LangGraph):

   • Coordinates state across three specialized nodes: recall_memory_node -> generate_copy_node -> brand_critique_node.
   • Ensures structured JSON generation and strict schema validation before UI rendering.

2. Persistent Assistant Memory (Backboard):

   • Stores brand personality profiles, forbidden vocabulary, and signature taglines inside a dedicated assistant workspace.
   • Eliminates the need to repeatedly paste guidelines into every prompt.

3. High-Speed Open-Weight Inference (Groq):

   • Powers the generation node using open-weight models (openai/gpt-oss-120b / open-weight LLMs) to return structured multi-channel copy within ~2–3 seconds.

4. Automated Guardrail Critique Node:

   • Compares drafted copy against forbidden negative constraints pulled from memory (flagging and rewriting buzzwords like "cheap" or "hurry").

5. Audio Synthesis (ElevenLabs):

   • Converts the generated Instagram Reel script into high-fidelity voice narration directly within the Streamlit workspace.

6. Web Application & Deployment (Streamlit + Render):

   • Streamlit frontend deployed to Render using a declarative render.yaml blueprint with environment variable isolation.

---

🤖 Why Does Open Innovation Matter?

Open innovation is what makes this project viable as an independent side hustle:

1. Vendor Independence & Portability: Building on LangGraph and open-weight model interfaces means the core business logic is not locked into a closed LLM provider. If pricing changes or a superior open-weight model drops, switching requires changing only an environment variable.
2. Data Sovereignty for Small Brands: Proprietary brand identity rules, unpublished product specs, and niche taglines remain isolated within private agent memory and open orchestration, rather than being fed into opaque public model training corpora.
3. Zero-Cost Baseline Economics: By combining open-weight inference with modern developer free tiers (Render, Backboard, Sentry), solo founders can deploy and validate a production-grade AI product without incurring high upfront infrastructure costs.

---

🏆 Prize Categories

I am entering the following prize categories:

1. Best Use of Render ($200 - Featured Category)

• Deployment Blueprint: MicroBrand Voice Engine is hosted as a web service on Render using declarative Infrastructure-as-Code (render.yaml).
• Production Configuration: Render handles automatic builds on Git pushes (pip install -r requirements.txt), dynamically binds Streamlit to $PORT, manages encrypted secrets (BACKBOARD_API_KEY, SENTRY_DSN, GROQ_API_KEY, ELEVENLABS_API_KEY), and serves the live app at https://microbrand-engine.onrender.com/.

2. Best Use of Sentry Agent Tracing ($100 - Partner Category)

• Full Observability: Integrated sentry-sdk to track every LangGraph state machine execution, agent tool call, and LLM transaction.

• What the Telemetry Revealed:

  • Agent Runs & Root Traces: Sentry automatically indexed LangGraph under Agents / Trace Root, measuring root durations across real runs (3.39s, 4.01s, 5.57s, and 6.80s).
  • Token & Cost Tracking: Live token tracking captured exact consumption per run (649, 745, 821, and 827 tokens) alongside cost estimation graphs for openai/gpt-oss-120b.
  • Real Production Debugging: The trace log captured an event-loop failure (ea0fae1f — 21.00ms execution with an error flag). Sentry's trace timeline pinpointed an httpx connection pool conflict in Streamlit, allowing us to isolate asynchronous tasks into dedicated thread runners.

3. Best Use of Backboard ($100 - Partner Category)

• Utilized Backboard's assistant memory to give the agent long-term persistence. Brand tone guidelines, negative keywords, and signature slogans persist across sessions and are recalled dynamically by the recall_memory node.

4. Best Use of ElevenLabs ($100 - Partner Category)

• Integrated the ElevenLabs API to generate realistic speech previews for social media copy, letting makers audition their Instagram Reel audio hooks directly inside the interface.

---

Built with ❤️ during Hacktoberfest 2026.