I’ve Deployed AI Inside Real Companies. These 4 Things Matter ( From Ex Palantir Engineer)

Summary

Anu Sharma argues that many AI portfolio projects look like simple prompt-and-response demos, making it difficult for recruiters to judge whether their creators can build usable products. She recommends demonstrating practical implementation skills through four features.

1. Build an AI agent, not just an AI app

An app may simply send a prompt to an LLM and display its answer. An agent should use tools, call APIs, retrieve information, make decisions, and take actions.

Examples include researching a company to personalize a cover letter, detecting data anomalies and notifying Slack, or querying a database and returning structured results.

2. Add retrieval-augmented generation (RAG)

RAG lets a system answer questions using company-specific, recent, or otherwise unavailable information. A basic workflow is to:

1. Split documents into chunks.
2. Embed the chunks and store them in a vector database.
3. Retrieve relevant passages for each query.
4. Generate an answer grounded in those passages.

Pinecone, Weaviate, and FAISS are mentioned as possible vector-store options. Answers should include links to their sources.

3. Provide a real-time chat interface

A product-like chat should support ongoing conversations, conversation history, and potentially multiple users sharing a session. Building this infrastructure from scratch can involve WebSockets, message queues, state management, reconnection, and reliable delivery.

The video promotes Nexconn—rendered inconsistently in the subtitles as “Nexson” or “Nexon”—a chat API and SDK with real-time chat, voice, and video capabilities, plus UI kits for web and mobile platforms. The speaker says its free plan includes full functionality for up to 10,000 monthly active users, requires no credit card, and includes multi-channel push notifications. The video also says the SDK works with Cursor, Claude, and Codex.

4. Put a live demo video at the top of the README

Recruiters may not clone a repository or follow multiple links, so the video recommends showing the project working directly. A proposed 60–90-second demo could follow this outline:

• First 10 seconds: Explain the problem and intended users.
• Next 20 seconds: Show the agent making a tool call or taking an action.
• Next 15 seconds: Demonstrate RAG answering a question using real data.
• Then: Show the real-time chat experience.
• Finish: Show the deployed URL loading in a browser.

Broader Recommendation

Build one complete, functional project rather than several generic tutorial projects. The goal is not simply to impress with novelty, but to make it easy for reviewers to see that the developer can build and deploy an AI product for real users.

Guides and Recommendations

• Build an agent that uses tools or APIs.
• Add a RAG pipeline using document chunking, embeddings, and vector search.
• Implement a persistent, interactive chat interface.
• Record a short working demo and link to it prominently in the README.
• The speaker’s Nexconn promotion says the free tier can be secured by signing up, integrating the SDK, and sending one message by December 31, 2026.

Main Speaker and Source

Anu Sharma is an AI engineer and technical content creator who presents the portfolio guidance and promotes Nexconn’s chat API and SDK.