Cheng AI: An AI-Native New Grad Trying to Make Himself Interesting

Website: https://chengai-tianle.ai-builders.space/

1. To be or not to be, here’s an answer

I feel like I’m stuck in a deadlock while job hunting. I keep hearing voices like these:

“New Grad SDEs do not ‘not exist’ because they shouldn’t. They ‘don’t exist’ because yesterday’s New Grad job descriptions no longer make sense today. The old filtering system can’t correctly select people anymore, and the new system hasn’t been built yet. So New Grads have disappeared.”

“Referrals are meaningless. Unless an executive forwards your resume directly to HR from the top, it doesn’t matter.”

“Senior engineers with architectural thinking plus coding AI can finish the work quickly. There is no place for juniors.”

Here is what I say.

First, even without AI coding, New Grad resumes all look the same. Getting selected out of thousands is not a normal distribution problem. It is an outlier problem.

Second, people argue that experienced engineers plus AI can do in one hour what five juniors used to do in a week, therefore companies should stop hiring juniors. But that misses something: juniors plus AI can also do in one hour what five juniors used to do in a week, and they are much cheaper.

Third, hiring filters were never great to begin with. A resume is nowhere near enough to represent me. So how can someone conveniently understand me, and decide whether they should hire me? Talk to my AI twin. It knows everything.

2. Four stages of evolution for a personal website

Stage 1: Treat yourself as a product

After watching the Career Signaling video from “Kebiaodai,” I built my portfolio site, published my ideas and projects, and polished my GitHub. I started treating myself as a product and validating product market fit. It was a long-held idea, but I never acted on it until now. I wanted to put my internships, projects, technical stack, and my thinking in writing onto a site, and let the site replace the resume.

Stage 2: If I’m going to be screened by AI, I might as well build my own AI

I know this cannot be the end. In the company where I interned, the team with the highest AI usage was HR. Today, many recruiting teams likely use AI to sift through thousands of resumes. If that’s the reality, instead of spending hours crafting a resume format that still cannot represent me, only to be reduced to crude keyword matching, I would rather build an AI based on my own life. Why not just talk to my AI? The pipeline is effectively the same.

I can also add a feature called JD Match. You paste a JD (Job Description), and Cheng AI retrieves from my history and work to estimate how well I match the role.

Stage 3: The website itself becomes my brand

If the website and the AI are well built and well deployed, and the quality is genuinely high, then the site itself demonstrates my engineering ability and my product and business thinking. The website becomes a living proof of competence, not just a container of claims.

Stage 4: A version of me that can be directly interviewed

My current stage goal is to build a version of myself that can be interviewed directly. Ask it technical questions, experience questions, or behavioral questions, and it is effectively the same as interviewing me.

3. Implementation

3.1 Writing and deployment

After dozens of rounds of conversations with different AIs, I pasted my long context into my own Abstraction AI (https://abstractai.ai-builders.space/) to generate implementation-ready specification documents. Then I fed the docs, BuilderSpace’s API and MCP, and a long prompt into Claude Opus 4.5 inside Augment Code, and it started building. Very quickly, the frontend, backend, Supabase, and deployment were done, and the first version shipped.

3.2 Iteration: Programming as training

I iterated using Codex and GPT-5.2 with extremely high reasoning. During iteration, I brought the mindset of training deep learning models into how I direct AI to write code: compare output to ground truth, compute loss, backpropagate, and do gradient descent.

Concretely, I asked Codex to read our long conversations and the specification documents, and infer what answering style and instruction set Cheng AI should follow. Then Codex talked to Cheng AI, evaluated how far the answers were from what we wanted, analyzed why the gap existed, and decided what to adjust: the prompt, the CRUD logic, or a specific detail in the RAG pipeline. Then it implemented the change.

This resembles an iteration or epoch in model training. Codex then talked to Cheng AI again, repeated the same evaluation loop, and kept iterating until there was nothing left to improve. This is similar to RLAIF (Reinforcement Learning from AI Feedback).

When iteration reached a point where the existing documents could not guide further improvements, and new requirements emerged, I used the same philosophy. For a new requirement, the loop became conversation, iteration, conversation, iteration, until the requirement was satisfied. This resembles RLVR (Reinforcement Learning with Verifiable Rewards).

4. Results

First, I got what I wanted. It is fast, flexible, and able to retrieve the information I want based on the question. In many cases, it answers better than I do. Technically, it uses hybrid retrieval (vector search plus full-text search) and fuses results via RRF (Reciprocal Rank Fusion) to improve retrieval accuracy.

Second, the system is complete, technically solid, and scalable. It showcases real engineering depth. Next.js 16, Supabase (PostgreSQL plus pgvector), Gemini 2.5 Pro, OpenAI embeddings. This is a real, usable RAG system.

Third, if you want to hire me, it is convenient. Talk to it. Give it your JD.

5. The future

Human attention is extremely limited compared to AI. Hiring is difficult and high-risk, and understanding a candidate costs a lot of tokens. So a likely future is this: the employer’s AI and the candidate’s AI will talk deeply, and discover the truly best-fit candidate.

Instead of interviewing me, interview my AI. It understands me and can retrieve the correct information, so the two of us will answer similarly anyway. My projects are also written with AI. The boundary between AI and humans is becoming increasingly blurry, and in mind, and eventually even physically, the merge may become irreversible.

6. Do you want your own AI?