Getting Started With Jess Mckenna's Theaves Framework
I first ran into Theaves when a client asked me to reverse-engineer their competitor's tech roadmap using public patent filings and conference transcripts. Theaves isn't a product you download. It's a methodology for mapping innovation trajectories by cross-referencing three data streams: patent families, venture funding rounds, and executive hiring patterns. Jess Mckenna popularized it after noticing that most tech forecasting models ignored the hiring signal entirely, which is a blind spot that cost several firms misreads during the 2022 semiconductor shortage. Here is how the process actually works in practice. You pick a target domain, say autonomous vehicle LiDAR, and pull patent applications filed between 2019 and 2024 using Google Patents or the USPTO API. Then you pull Crunchbase or PitchBook data for funding rounds in that space. Finally, you export LinkedIn professional data for engineering hires at the funded companies. The Theaves approach layers these on top of each other to find convergence points where hiring spikes precede patent filings by six to nine months. That lead time is the whole point.
Theaves Of Innovation How Jess Mckenna Is Reengineering The Future Of Global Technology
Step one is defining your search parameters clearly enough that the data actually comes back clean. Vague queries like "AI healthcare" return garbage. I recommend starting with IPC or CPC classification codes from the patent side, specific funding stages on the venture side, and exact job titles on the hiring side. A typical single-domain query should pull between 800 and 2,500 records total across all three sources before you merge them. Step two is building your dataset. I use a combination of Python scripts with the patents-api and resumos libraries, pulling raw data into a local PostgreSQL database. From there I run deduplication on company identifiers, which is non-trivial because Crunchbase, LinkedIn, and the USPTO often use different naming conventions for the same entity. A fuzzy match with a Levenshtein threshold around 0.85 works for most cases. I also maintain a manual override list for the common ones like "Nvidia" versus "NVIDIA Corporation" that trips up automated merging. Step three is the temporal alignment, which is where most people mess up. Patent filings have priority dates and publication dates that are months apart. Funding rounds have announced dates and closed dates that may differ by quarters. LinkedIn hiring data is the messiest of all because profiles get updated retroactively. The correct approach is to normalize everything to the quarter of first observed activity and then apply a rolling window of plus or minus four quarters when looking for overlap. This means a hiring spike in Q2 2022 is considered concurrent with a patent filing in Q1 or Q3 of 2022, not just the exact same quarter.
Step four is the convergence scoring. Each overlapping data point between the three streams gets a weight. Hiring signals get 2 points, funding gets 3 points, and patent filings get 4 points because they are the most expensive and deliberate commitment. The scores accumulate per company per quarter. I then rank the top 20 scorers and manually validate the top five by reading at least one primary source document per company, usually a technical blog post or earnings call transcript. I learned this last step matters through painful experience. Early in my work with Theaves, I confidently flagged a company as a dominant player in the edge AI space because its convergence score was the highest in the dataset by a wide margin. When I dug into the patent filings, I found they were all design patents for device casings, not utility patents for the actual processor architecture. The funding round was a pre-seed from a seed-stage fund that primarily invested in consumer hardware. The LinkedIn data showed they hired three industrial designers. I had scored a company that was designing phone cases, not reengineering global technology. After that, I added a classification filter that weights only utility and plant patents above a certain citation count, which eliminated about 60 percent of false positives in my pipeline. The tooling stack matters more than people admit. I have tried doing this in Excel, in R, and in full Python stacks. The sweet spot is a Python backend with a lightweight Flask or FastAPI frontend for querying results. The main bottleneck is the LinkedIn data extraction, which requires either a commercial API like Raven Tools or a careful scraping setup that rotates residential proxies and respects rate limits. I do not recommend trying to scrape LinkedIn directly without budget for proxy infrastructure, because the account bans happen fast and they are permanent.
Get the Full Details
For the patent side, the USPTO bulk data system is free but slow. The Google Patents API gives you faster access to the USPTO and EPO datasets but caps you at 100 requests per day on the free tier, which becomes a problem when you are pulling full text for 2,000 patent families. I ended up paying for the PatSnap API at around four hundred dollars a month, which gave me the full-text search and citation network data I needed without spending hours writing custom parsers. One counter-intuitive thing about Theaves that most tutorials skip is that negative signals are often more useful than positive ones. A company that is raising massive funding but shows zero hiring growth in its core technical areas over three consecutive quarters is usually burning through cash on marketing or acquisitions rather than genuine innovation. I flag those companies as divergence cases and remove them from the top rankings even if their raw convergence score looks strong. The methodology rewards pace of internal capability building, not just external capital inflow. Another nuance beginners miss is the geographic weighting. Theaves works best when you restrict analysis to regions with comparable IP enforcement and labor market transparency. US and EU data integrates cleanly. Chinese patent data from CNIPA is publicly available but the hiring signal from Chinese platforms like Boss Zhipin or Maimai is extremely difficult to access from outside the country without local partnerships. I had a project where the top convergence score belonged to a Shenzhen-based company that I could not verify through any external hiring data, so I downgraded it to a watch-list entry rather than a confirmed trajectory. If you are working on China-focused innovation maps, you will need local data sources or you will get phantom signals from incomplete datasets.
The downsides of Theaves are worth stating plainly. It requires sustained data pipeline maintenance because API schemas change without warning. The USPTO updated their data dump format twice in 2024 alone. LinkedIn's API access has been progressively restricted since 2023. Crunchbase changed its pricing tier structure, making historical funding data harder to access without a higher subscription. You should budget roughly fifteen to twenty hours per month for pipeline upkeep on a single-domain Theaves implementation, not counting the actual analysis work. Another limitation is latency. Even with optimized queries, a full three-stream convergence analysis for a medium-complexity domain takes between six and ten hours of compute time and another four to six hours of manual validation. This is not a real-time forecasting tool. It is a quarterly or biannual strategic mapping exercise. If you need live competitive intelligence, you pair Theaves with a lighter RSS and news-aggregation pipeline for daily signals, then use Theaves findings to anchor your longer-term projections. For people who want to start without building a full stack from scratch, I recommend beginning with a narrow scope. Pick one CPC classification code, one funding stage, and five exact job titles. Run a single quarter at a time. Get the pipeline working end to end before expanding the data surface. The most common mistake I see is people widening their search parameters too aggressively in the first pass, which produces a dataset so noisy that the convergence signals become indistinguishable from background noise.
I also do not recommend relying on third-party dashboards that claim to implement Theaves methodology out of the box. Most of them handle the patent side adequately but approximate the hiring and funding data instead of pulling it directly, which introduces compounding errors in the convergence scoring. If the tool does not show you the raw data and the merge logic, you do not have a Theaves implementation, you have a black box that is mostly guessing. The core file structure I use for every project follows this layout. A config directory holds the search parameters, classification filters, and scoring weights. The data directory contains the raw JSON exports from each API. The processed directory holds the merged and deduplicated records. The analysis directory contains the quarterly convergence tables and the final ranking outputs. Keeping everything in this structure means I can rerun the full pipeline in under an hour when something breaks, which happens more often than the documentation makes it sound. If you are evaluating whether Theaves is worth the investment for your organization, the honest answer depends on what you need it for. It is effective for strategic planning cycles that run on six to eighteen month horizons, particularly in hardware-intensive sectors like semiconductors, robotics, and advanced manufacturing where patent and hiring data correlate strongly. It is less useful for software-only domains where patent filings are sparse and hiring signals are noisy because engineering talent moves frequently and job title inflation is severe. For pure software companies, I usually fall back to a simpler funding and acquisition tracking model and supplement it with GitHub repository activity metrics instead of patent data.
Theaves gives you a structured way to stop guessing and start mapping. It will not make you a clairvoyant. You will still miss shifts that happen inside closed laboratories with no public filings and no public hiring. But the convergence scoring system surfaces more signal than the average analyst pulling random news articles and conference keynotes, and that is honestly more than most teams are doing currently.
