# Humata TechStack Analysis

## Tools found with [`openapi-devtools` Chrome Extension](https://github.com/AndrewWalsh/openapi-devtools)

- https://nextjs.org/
- https://vercel.com/
- https://vercel.com/domains (https://www.whois.com/whois/humata.ai)
- https://stripe.com/
- https://supabase.com/pricing
- They use it for all auth, database and object storage
- They send a verification email:
- `Confirm your signup
Please confirm your email address by opening the link: https://benzfodquatjcaxqhbmk.supabase.co/auth/v1/verify?token=XXXXXXX&type=signup&redirect_to=https://app.humata.ai`
- The subdomain indicates they use the managed service
- https://launchdarkly.com/
- DevOps tools, A/B tests, Feature flags, Session recording, ...
- http://app.posthog.com
- Similar to launchdarkly
- http://api-iam.intercom.io
- Not sure for what exactly
- `iam` means `Identity and Access Management`, but they already use Supabase for that.

## Hacks

- I circumvented their **free plan PDF 60 pages limit** by merging 4 pages into one using http://www.pdfdu.com/pdf-pages-merge.aspx and was able to upload the whole [EU AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A52021PC0206) `PROPOSAL` (108 pages -> 27 pages) and `ANNEXES` (17 pages -> 5 pages).

## Insights

### LLM model used

Their analytics tool https://clientstream.launchdarkly.com/ made a request and exposed their used LLM model:

patch	{"key":"open-ai-model","value":"gpt-3.5-turbo"}	

### Cabability evaluation

- A tricky question like **`give me only all the titles from the proposal document`** is hard because it would need to analyze all pages contents.
- Instead it returned wrong output (some other `titles` mentioned in the document) that indicates Humata uses similarity search and does not have secret sauce algorithms.
- Their internal steps are likely basic RAG:
1. Generate search queries from user question
2. Search backend using preprocessed documents
3. Generate chat response using augmented content


### Is Humata capable?

- No, it is very basic and is unusable for complex legal document work as proclaimed by their [blog post](https://www.humata.ai/blog/how-to-use-ai-for-legal-documents).
- Their service is a basic RAG with a nice website, although they don't even render messages as clean Markdown, thus the chat experience is very bad.
- The only nice addition is the document view and highlighting on the right.
- **The next screenshot underlines the unusefullness of the product.**


## Conclusion

- Robin got very excited about Humata, but it turns out, the general case as with most AI startups persists: "Their marketing is better then the product."
- Robin wonders, once again, how easy it is for startups to get **$3.5M** fundraising from Google Ventures without any exceptional technology.
- We can build a product on par with Humata.
- One problem ComplyAI has is our current website design, branding work and vision **is not cohesive enough**. _We need better design._