Skip to content

Commit

Permalink
Ansa edits on case study (#1080)
Browse files Browse the repository at this point in the history 
* Ansa edits on case stuyd

* small fixes

---------

Co-authored-by: Tom Gotsman <tomgotsman@Toms-MacBook-Pro.local>
  • Loading branch information
@rsullivan05
rsullivan05 and Tom Gotsman authored Nov 13, 2024
1 parent 49b2ddc commit f2467e2
Showing 1 changed file with 49 additions and 78 deletions.
127 changes: 49 additions & 78 deletions case-studies/ansa.md
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
---
company: Ansa
description: "Why Ansa chose Reflex over Anvil for their data analysis workflow automations"
description: "Why Ansa chose Reflex over a no-code/low-code framework for their workflow automations"
domain: "https://www.ansa.co"
founded: "New York, 2021"
investors: ""
Expand All @@ -11,11 +11,11 @@ stats: [
},
{
"metric": "Core company workflows optimized",
"value": "6"
"value": "8"
},
{
"metric": "Hours of manual work saved a month",
"value": "650+"
"value": "100+"
}
]
meta: [
Expand All @@ -39,160 +39,131 @@ meta: [
---


Meet [Ansa](www.ansa.co), a venture capital firm based in New York City that invests in companies from Series A to C. They have invested in companies such as Coinbase, Rippling, Glassdoor and Godaddy.
Meet [Ansa](www.ansa.co), a venture capital firm based in New York City that invests in companies from Series A to C. They have invested in companies like Defense Unicorns, Bland, Gradient, and Selector and previously backed Crowdstrike, Coinbase, and SurveyMonkey.

Ryan Sullivan is an investor and manages the tech and data science team at Ansa. He works to help find the right companies to invest in and improve Ansa's automated outbound strategies.
Ryan Sullivan is an investor and manages the engineering and data science team at Ansa. He finds and supports new investments and builds Ansa's data-driven sourcing strategy.


## Ansa's overwhelming data analysis challenge
## Ansa's data analysis challenge

As a VC, Ansa must filter through tens of thousands of companies a month to find the best possible investment deals. Their sourcing is very outbound driven and previously they could never review all the companies they wanted to, meaning they were missing potentially great deals.
Ansa has an investable universe of 10s of thousands of companies and they need to make sure they are spending time with the right companies at the right time. Their sourcing is thesis driven, so they need to both quickly find all companies in a theme of interest and track opportunities across the broader market. With a lean investment team, they need to leverage software and data to review and track all these opportunities.

Ryan and his team's goal was to automate as much as they could on this company analysis workflow. This included helping their smaller investment team be more efficient and effective in finding interesting companies and reaching out to them. They also began using Data Science and Machine Learning to help improve the sourcing of companies and to increase automated outbound.
Ryan and his team's goal was to automate and augment as much as they could on this company sourcing and review workflow. This included helping their lean investment team be more efficient and effective in finding interesting companies and reaching out to them. They also began using Data Science and Machine Learning to help surface more relevant companies leverage proprietary and third party data sources.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
We're frankly a smaller team and there's a lot of opportunities out there, so we're trying to automate as much as we can on the workflow side to help our smaller team be efficient and cover as many companies as possible, researching them and scaling out our outbound.
- role: Investor and Head of Data
We have a lean investment team and there's a lot of opportunities out there, so we're trying to automate as much as we can on the workflow side to help our team be efficient and research, review, and reach out to as many companies as possible.
```

They wanted in-built tools to give them an edge as a small team, such as an in-house scoring model to help flag important companies.
They wanted custom tools to give them an edge, such as an in-house scoring model to help flag important companies.


## How Ansa hit the limits of the Anvil framework
## How Ansa hit the limits of a low-code python framework

Ryan and his team wanted to build a web interface for their team to use to automate their currently very manual workflows. They wanted a pure python solution, as this was the language that most people in the team were most familiar with.
Ryan and his team wanted to build a web interface so the broader team could run these automations to automate their manual workflows. They wanted a pure python solution as this was the language the team was most familiar with.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
We don't have a full engineering team, so to build a full web app from scratch seemed like a lot for me to do. In addition, we don't have experience on the JavaScript side.
- role: Investor and Head of Data
We don't have a full engineering team, so to build a full web app from scratch seemed like a lot to manage. In addition, our team is mostly data engineers / analysts so we are far more comfortable with Python than JavaScript.
```

The team previously built an [Anvil](https://anvil.works) app, a near no code framework. They didn't like the aesthetic and wanted to use more modern looking React components.
The team previously built on an all-python, low-code / no code framework. They didn't like the aesthetic and wanted to use more modern looking React components.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
It’s an older framework and the components didn't really look that good. We wanted to use react components and just make it look a little bit more modern.
- role: Investor and Head of Data
It’s an older framework and the components didn't look that good. We wanted to use react components and just make it look a little bit more modern.
```

Their main concern with Anvil though was that they didn't want to outgrow a near no code framework, as they wanted to build their app for the long term.
Their main concern though was that they didn't want to outgrow a near no code framework, as they wanted to build their app for the long term.


```md quote
- name: Ryan
- role: Investor and Head of Data Science
We don't want to run into a situation where this year or next year, we want to add more functionality that Anvil doesn't have and we're not able to integrate it. We know we're building this for the long term and we want to have flexibility and not outgrow it.
- role: Investor and Head of Data
We don't want to run into a situation where this year or next year, we want to add more functionality that this low code framework doesn't have and we're not able to integrate it. Additionally, the rate of improvement and development velocity from the Reflex team gave us confidence that their offering would continue to improve over time. We're building this for the long term and we want to make sure we both have the flexibility to not outgrow it and are working with the best out there.
```

In addition, there were particular technologies like LLMs and Vector Databases that Ryan and the team knew at some point they would want to integrate into the app. This would be extremely difficult if not impossible with low/no code frameworks like Anvil.
In addition, there were particular technologies like LLMs and Vector Databases that Ryan and the team knew at some point they would want to integrate into the app. It would be extremely difficult if not impossible to keep up with these latest innovations with low/no code frameworks.


```md quote
- name: Ryan
- role: Investor and Head of Data Science
I started to feel that with Anvil I didn’t know if we can do everything. If they don't have a component for it, I wouldn’t know how to integrate it. For example some of the newer stuff we do with vector databases and embeddings or LLMs would be harder to do in Anvil as we would be stuck with the integration that Anvil have.
- role: Investor and Head of Data
I started to feel that with this framework I didn’t know if they could keep up with the pace of new developments with LLMs. They abstract a lot of the backend, so it's difficult to install third party libaries and you don't have full control over the database. For example some of the newer stuff we do with vector databases, embeddings models, or LLMs would be harder to do with this framework as we'd have to move off their native database.
```


## From manual work to automation with Reflex

Ansa switched from Anvil to Reflex because they could build both a web app for the long term that they would not outgrow and they required no JavaScript experience.
Ansa switched to Reflex so they could build an app for the long term and accommodate all the latest innovations in LLM development without needing any JavaScript.

They currently have an app with 6 different core company workflow automations, several of which we will discuss in this case study.
They currently have an app with 8 different core company workflow automations, several of which we will discuss in this case study.

The main challenges that Ansa face are how to figure out what companies to spend their time investigating further to potentially invest in, and then finding the relevant information to reach out to these companies.
The main challenges that Ansa faces are one, figuring out what companies, out of the 10s of thousands within their investment mandate, they should be investigating further, and two, automating all the manual data collection and work required to reach out.


### Creating natural language company search

The first automated workflow they built, using a combination of OpenAI and Langchain, introduced natural language to SQL searches for a database. This allows employees to more easily search for companies based on semantic similarities.


### In-house company ranking algorithm

They built another automated workflow to take this long list of companies and rank them. They run this workflow every week. It takes a list of companies, sends them to their API, and there each company is scored by running their custom AI model in Databricks.

These scores are returned back and emailed to the user as a CSV. This AI model, they built in-house, is used to spot factors in companies they believe will lead to successful investments for the fund.


### Increasing understanding of what a company does using vector embeddings

Once they have a list of interesting companies to explore further the team created an automated workflow to quickly get a deeper understanding of what the companies do.

This workflow scrapes company data from the internet and creates vector embeddings using an LLM to allow their employees to do a vector search through the companies to quickly understand what that company does.

The first automated workflow they built, using a combination of OpenAI, Langchain, and Chroma, introduced vector and natural language searches over their database. This allows employees to combine quantitative and strict filtering with an understanding of the companies product offering through vector similarity. For example, an employee can type "Carbon accounting software companies with a CEO in NYC that score over 60" and receive a curated list of companies that fit that description.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
- role: Investor and Head of Data
We use LLMs to help navigate through the companies site and find different details. For example the customer page for one website, may be different from another. The LLM then summarizes all that data and creates embeddings on them and then we use that for the searches. The LLMs help us normalize across different companies, even if pages are named differently, so we can easily search through all of them and figure out what the company does.
```

### In-house company scoring algorithm

The next automated workflow takes this list of companies and scores them. With private companies, there is far less data available to assess fit than with public companies, so they rely on alternative data to power a scoring algorithm that assesses the probability a given company is a fit for their investment workflow. They proactively score ~15K companies and display them in Reflex, and also built another automated workflow to score ad-hoc lists of companies. This workflow can take in a list in any format and send the identifiers to their API where they are scored by their custom ML model hosted in Databricks.

The scored data is then displayed to the user in Reflex and emailed to the user as a CSV. This ML model is trained on a labeled dataset they have curated over years, and spots combinations of factors that they believe will lead to successful investments for the fund.


### Improving email outbound

Finally, when their team has a short list of companies they like, they built a fourth workflow to automate the extraction of the relevant information to reach out to these companies. Ryan runs us through this final workflow in his quote below.
Finally, when their team has a short list of companies that fit within an investment, they built a third workflow to automate the extraction of the relevant information to reach out to these companies. Ryan runs us through this final workflow in his quote below.


```md quote
- name: Ryan
- role: Investor and Head of Data Science
Let’s say we have 20 companies that we want to email. How do you actually reach out and email those companies? We launch a script, that runs through a Reflex background event, that'll go through each company, check the CRM ownership, it'll tag it, fill out relevant fields and find the best person to reach out to. A lot of times, especially with early stage companies, it's not always clear who the founder or CEO is or what their email is. So this workflow will find their email, test all that, and then it will go to our email engagement and tracking tool and add it there and make sure everything is relevant. Before we would do this all manually. Now with this new workflow built in Reflex it is as easy as click and it’s done.
- role: Investor and Head of Data
Let’s say we have 30 companies that we want to email. How can you efficiently send a custom note to each of these companies and track it properly? We launch a script, that runs through a Reflex background event, that'll go through each company, check the CRM ownership, fill out relevant fields and find the best person to reach out to. A lot of times, especially with early stage companies, data is missing or partially complete. So this workflow will leverage LLMs throughout the process to handle fuzzy matching and make contextual decisions, as well as proactively summarize company content, news, and relevant Ansa content to help support the email writing. Before we would do this all manually, now with this new workflow in Reflex, we've taken what was once 30+ clicks across 5 different apps and made it 5x faster with 2 clicks across 2 apps.
```


All these different workflows, are now built into a single Reflex app. It makes it extremely easy for anyone on the team to run any of these workflows and move data from one to another with a few clicks.
All these different workflows are now built into a single Reflex app. It makes it extremely easy for anyone on the team to run any of these workflows and leverage LLM-powered automation with a few clicks.

Throughout building this Reflex app, Ryan used:

- Supabase database to store all their data

- LLM tools like OpenAI, Tavily, Browserbase, Langchain, and Chroma

- Google Auth login component for Ansa employees to log in

- Download Functionality
- AG Grid Table Component

- Table Component
- Download Functionality


Overall Ansa found that with Reflex, as everything is in pure python, they were able to integrate everything they wanted and knew they always could, which was their major concern with Anvil.
Overall Ansa found that with Reflex, as everything is in pure python, they were able to integrate everything they wanted and knew they always could incorporate new tech, which was a concern with their previous framework.


```md quote
- name: Ryan
- role: Investor and Head of Data Science
Reflex was a better fit than Anvil. Given that it's just python code, I'm always comfortable that we'll be able use different tools and to figure out how to make it work with Reflex versus being stuck with the integrations that Anvil have.
- role: Investor and Head of Data
Reflex was a better fit overall. Given that it's just python code, I'm always comfortable that we'll be able use different tools and to figure out how to make it work with Reflex versus being stuck with the integrations that our old solution had.
```


## What Ansa gained from using Reflex

The app that Ryan and his team created, which contains 6 different automated workflows, is now a central dependency for Ansa to source potential companies and analyze them.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
75% of our team uses the app on a weekly basis.
```

They process and score 16,000 companies on a monthly basis using their in-house company ranking workflow and show all that data in the Reflex app.

They then take around 4000 of the highest scoring companies a month and run them through the vector embedding workflow to research the companies.
The app that Ryan and his team created, which contains 8 different automated workflows, is now a central dependency for Ansa to source potential companies and analyze them.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
Our benchmark is that the automated workflows for researching the company and adding it into the outreach cadence saves 10 minutes per company versus doing it manually.
- role: Investor and Head of Data
75% of our team uses the app on a weekly basis and we estimate we're saving over ~100 team hours per month.
```

Overall this is saving over 650 hours a month for his team.

Finally, the combination of the automated workflows is not only saving the team time, it is also bringing better deals to the table that may have otherwise been overlooked.

```md quote
- name: Ryan
- role: Investor and Head of Data Science
Then there's also opportunities that get flagged and displayed that we probably otherwise would have missed.
```

0 comments on commit f2467e2

Please sign in to comment.