Hypothesis
As a data scientist seeking a job in AI, I often wonder: how do I figure out what kind of AI work a company is actually doing? Job postings only give you so much, and they can be vague or lag behind the company's true focus. Which led me to think... is there a way to understand what a company is actually investing in when it comes to AI, and then translate that into information that can guide my career decisions?
Here's a nerdy fun fact: companies file patents on what kind of AI work they're investing in. And those records are public. You can look up what AI subfields Microsoft or Google is pouring resources into, how that focus has shifted over time, and how concentrated or spread out their AI strategy is.
So the main question is: if you're trying to break into AI at a specific company, how do you figure out what they actually need? Job postings tell part of the story, but how can we use patent data to get a deeper understanding of a company's AI focus and translate that into something actionable?
That became the core of Agentic Evolution: take patent data, financial indicators, and labor-market signals, connect them together, and build a tool that gives you career intelligence you can actually act on. We built this for DTSC 5082 at UNT.
Finding the Data
So we needed three things: what companies are patenting, how they're doing financially, and what skills the job market actually wants. Here's what we ended up using:
USPTO AI Patent Dataset (AIPD) has AI patent activity broken down by technical subfields (computer vision, NLP, hardware, planning/control, etc.), all mapped to the companies that filed them. This is how we built each company's "AI innovation profile."
Compustat gave us firm-level financial data like revenue, market cap, sector, and growth rates. We needed this to check whether a company's AI investment was actually translating into market performance.
LinkedIn Skills Data (from HuggingFace) covered job-market skill trends. This was our bridge between "what a company patents" and "what a company hires for."
Getting these three datasets to play nicely together was honestly harder than the modeling. Entity resolution alone... matching company names that are formatted differently across each dataset ("MICROSOFT CORP" vs. "Microsoft Corporation" vs. "MSFT") ate up more time than I'd like to admit.
How It All Connects
Once we had the data figured out, the pipeline goes from a company name to a full career intelligence report. Here's the flow:
- User picks a company from a dropdown in the Gradio app
- Entity resolution matches it to the patent dataset
- Pull the patent profile + financial features
- Run through a StandardScaler preprocessor
- Random Forest Regression predicts revenue growth
- Random Forest Classification assigns a growth tier (High/Moderate/Low)
- Patent subfields get mapped to in-demand job skills
- Everything gets passed to Gemini 2.5 Flash for a RAG-powered career report
- Gradio UI displays charts, skills, market context, and the report
The Models
We ended up building two Random Forest models. The first one, Revenue Growth Regression, predicts expected growth based on AI patent characteristics like subfield breadth, concentration, and filing velocity, combined with financial controls. Basically asking: do a company's AI investment patterns correlate with stronger growth?
The second is a Growth Tier Classification that buckets companies into High/Moderate/Low growth. Honestly this ended up being more useful for the career intelligence layer than a raw percentage. Easier to act on.
One thing that was cool to see: we computed an AI Intensity Score (0 to 1) capturing how much of a company's innovation output is AI-related. Microsoft scored 0.56, meaning over half its patent activity is AI-focused. Some companies you'd think of as "AI companies" scored lower than I expected because they patent heavily in other areas too. That surprised me.
The Career Intelligence Report
This is the part that made the project feel like it actually does something. Once we have a company's AI profile, we pass it as context to Gemini 2.5 Flash through a RAG setup. The model generates a structured report that answers:
- What kind of AI work is this company actually doing? Specific areas backed by patent evidence, not buzzwords.
- Top 5 skills to prioritize? Mapped from patent subfields to concrete technical competencies.
- What role titles are they likely hiring for? Based on patent focus crossed with job posting patterns.
- One specific career advice tip. Tailored to the company's actual AI strategy.
For Microsoft, the system picks up that 55.9% of their AI patents are in hardware. So instead of generic "learn Python" advice, it points to CUDA, ONNX, quantization, pruning, and suggests contributing to open-source projects like ONNX Runtime. That's way more useful than anything I've gotten from a job posting.
I was actually surprised by how useful the generated reports were. Gemini doesn't know anything about Microsoft's patent portfolio on its own, but when you inject that data as context, the output gets really specific. The quality of the retrieval step directly shapes the quality of the generation. That was probably the biggest lesson from this project for me.
The Demo
The Gradio app shows four things when you select a company:
- AI Research Areas -- pie chart of patent portfolio by AI category
- Patent Activity Over Time -- line chart of AI investment growth
- Recommended Skills -- top skills based on research focus
- Research Focus Over Time -- how priorities have shifted
The whole analysis runs in a few seconds. It was satisfying to see everything come together into something you can actually click through and explore.
Team
This was a group project. My teammates: Akhil Sai Yalavarthi, Hoda Malak, and Sreekanth Taduru. We split work across three phases: EDA and data integration, feature engineering and modeling, then agent orchestration and deployment.
What I Learned
I had never worked with USPTO data before this. It reveals things about a company's technical direction that you just don't get from press releases or earnings calls. That was genuinely new for me.
The notebook-to-app gap is also very real. We had working models in notebooks, but getting them into a clean pipeline with proper artifact management (saving and loading scalers and models from pickle files) and a user-facing interface took just as much effort as the modeling itself.
The biggest takeaway was about RAG though. Gemini on its own can't tell you anything about a specific company's patent portfolio. But feed it that data as context and the output becomes targeted and useful. The retrieval step matters way more than I initially thought.
Also, skills mapping involves a lot of judgment calls. "Computer Vision" maps to OpenCV and PyTorch fairly cleanly. But "Planning/Control" could mean robotics, supply chain, or autonomous vehicles depending on the company. We handled this with context-aware mapping, but it's an area with room for improvement.
Tech Stack
- Data Sources: USPTO AIPD, Compustat, LinkedIn Job Postings (HuggingFace)
- Modeling: scikit-learn (Random Forest), StandardScaler
- LLM: Google Gemini 2.5 Flash (RAG pipeline)
- App: Gradio
- Data Processing: Pandas, NumPy
- Visualization: Matplotlib, Plotly
See It in Action
Built for DTSC 5082 at UNT | Group 8