The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Back to feedPost
The network for creativity
Join 1.25M professional creatives like you
Connect with clients, get discovered, and run your business 100% commission-free
Creatives on Contra have earned over $150M and we are just getting started
Trending
retrofuturism
Share your favorite designs from the past exploring the distant future. Or your own take on the aesthetic 🚀
halftone
Dots. Texture. Grain. Let's see your designs using the halftone technique.
Midjourney
Creatives are showing off what they're making with Midjourney.
colorgrade
Video editors sound off: How do you enhance the mood of your video in color grading?
Related posts
My core stack is less "generate pretty pictures" and more "build the pipeline that does the thinking":
n8n — the glue. Great for orchestrating multi-step agent workflows and connecting APIs without writing a service from scratch. Weakness: gets messy fast on complex branching logic, and debugging a broken node chain eats time.
RAG (usually with a vector DB + LangChain-style retrieval) — lets me ground LLM output in actual client data instead of hallucinated generalities. Strength: dramatically improves accuracy for domain-specific asks. Weakness: chunking/retrieval quality is finicky, and a bad embedding setup silently tanks results without throwing an error.
Claude/GPT for reasoning + code — my default for anything needing judgment: analysis writeups, SQL/Python generation, proposal drafting. Strength: fast iteration. Weakness: still needs a human check on numbers — confident-sounding ≠ correct.
Power BI / Tableau — where the AI output actually becomes something a client trusts. AI can shortcut the analysis, but the dashboard is what gets signed off on.
The honest lesson: the "AI" part is maybe 30% of the workflow. The rest is validation, data plumbing, and making the output legible to a non-technical stakeholder.
Silent failures are what I watch closest. I treat an answer as untrusted unless I can show exactly which source text the model read, because fluent output can hide a bad match.
This project demonstrates a Retrieval-Augmented Generation (RAG) architecture using open-source tools. It solves the "knowledge cutoff" problem of LLMs by giving them access to live tools.
How it works:
Intent Classification: The user's input is first processed by a local Llama 3 instance, which acts as a semantic router. It analyzes the prompt to determine if external data is required.
Tool Execution: If a search intent is detected, the Python backend executes a query against the SerpApi (Google Shopping) endpoint. Data Structuring: The raw JSON response from the API is parsed, cleaned, and structured into standardized product objects.
UI Rendering: The structured data is dynamically rendered into interactive product cards using Streamlit's component system, allowing users to view images and click direct purchase links without leaving the chat interface.
I designed and built an AI Web Research Agent using Python, Streamlit, OpenRouter-compatible APIs, DuckDuckGo search, Requests, and Trafilatura. The workflow creates targeted search queries, gathers and ranks results, reads available webpages, preserves source URLs, distinguishes fully fetched pages from search-result-only sources, and generates a structured Markdown report. I also created the interface, live activity feed, quick demo prompts, source-quality labels, verification-limit section, and downloadable report experience. The output is designed as a research draft that remains subject to human review.
