AI Software Development Company
Services
- Full-cycle AI Software Development
- Data Warehousing, ETL, Business Intelligence (BI)
- Advanced and Predictive Analytics
- MLOps, DevOps, and other software infrastructure development
Value Proposition
- AI Software Development - FASTER, CHEAPER, and BETTER
- Top notch in Healthcare, Life & Chemical Sciences
Recommended Providers
Focus
Portfolio
HTG Molecular, One.Skin Technologies LLC, Sweet TV, National Healthcare, ICS, Legendari, Ulm University Hospital, HTG Molecular Diagnostics, NephroSant
Predicting readmission risks
This project was delivered for a healthcare provider that makes patients successful in therapies through the individual attention of skilled nurses and moving to well-equipped small facilities. The provider guarantees effective care and fully covers insurance if patients are readmitted back to the hospital.
Business Goals
- Reduce the number of readmissions
- Upgrade care management operations
Challenge
- One readmission costs ~$20 000 for the healthcare provider. Some patients have 11 readmissions during a 90-day episode, totaling ~$220 000 loss.
- Nurses are overloaded and need to prioritize patient care. One nurse can have ~400 beds.
Solution
- 6 years of medical history records and machine learning tools were used to uncover reasons, predict and prevent readmissions.
- 3 groups of patients have been identified according to the readmission risk rate.
- Detected readmission reasons became the foundation for personalized health plans.
- Each group of patients gets a target health plan to reduce readmission risks.
- Each patient is tracked by predictive AI models and allocated to one of the risk cohorts in real-time.
- Nurses get push notifications about patient risks.
- Nurses can add patient notes taken into account by the AI service.
Impact
- In 90 days of the A/B testing period, the number of readmissions dropped by 32% in the experimental group nursed with the AI service, compared to the control ****group ****under ordinary care.
- Recalculated to the number of patients, the savings were ~$8 000 000 for the observation period (90 days).
Detection of opioids overprescription
This project was delivered for a pharmacy spending management company that serves long-term care and skilled nursing operators.
Business Goals
- Detect and prevent overprescription of opioids
Challenge
- The usage of opioids is associated with severe risks. These risks include misuse, addiction, overdoses, and death.
- According to the US Centers for Disease Control and Prevention (CDC): ”The number of drug overdose deaths increased by nearly 5% from 2018 to 2019 and has quadrupled since 1999. Over 70% of the 70,630 deaths in 2019 involved an opioid.” Source
- Detect doctors that are overprescribing opioids.
- Detect patients that are taking addictive substances on demand rather than in genuine need.
Solution
- Years of medical prescriptions history and a stack of machine learning tools for anomaly detection were used to detect doctors that prescribed and patients that got excessive amounts of opioids.
- Then, the detected cases of opioid misuse were analyzed by a team of experts. Such factors as patient age, gender, route of administration, dosage form, location, etc. were taken into account. As a result, the system of post-filtering for the anomaly detection algorithm was designed
- The anomaly detection model and guard rails bearing domain knowledge produced an effective tandem allowing reliable detection and preventing cases of opioid misuse.
Impact
- The service detected 367 cases of opioid overprescription in the sample of nearly 500 care settings and 12 200 people under the study. The cases are under closer scrutiny now.
Data Warehouse, ETL, Business Analytics
Customer is a streaming platform that offers a wide variety of TV shows, movies, documentaries, and more on thousands of internet-connected devices.
Solutions
Total Cost of Ownership (TCO) was carefully considered by comparing the performance and costs of alternative services offered by a number of cloud providers. The most advantageous combination of cost, quality, and sustainability was chosen to meet customer requirements.
Business requests were collected by interviewing numerous stakeholders. The requests served as a starting point for the BI dashboards, the DWH schema, and data pipelines design.
The multidimensional OLAP cube (facts, dimensions, etc.) was built and extended piece-by-piece revealing one business request after another. The cloud warehouse consolidated large amounts of data from disparate on-premises sources.
Convenient BI dashboards were designed and implemented to meet collected earlier business requests and their possible future variability. The end business users with no SQL skills could perform requests using straightforward BI dashboards, without DBA team involvement.
Auto-scalable ETL workflows were designed and implemented as the business requests and DWH schema had been established. The pipelines extract, transform, and load data from the source databases (on-premises) into the target DWH (on the cloud).
Backward historical consistency was achieved by introducing meta-tables and unique timestamps that indicated the time range when a source row or table was actual.
Impact
Using simple controls, business users can configure interactive dashboards and get required information immediately and independently, without the involvement of other departments.
DBAs are now free from writing the exact same SQLs every day. Finally, they’ve got time to clean up and refactor on-premises data sources. Without fear of breaking analytics - because the cloud DWH securely stores all past states of the originals.
Movie Recommender System
Customer is a streaming platform that offers a wide variety of TV shows, movies, documentaries, and more on thousands of internet-connected devices.
Business Goals
- Improve user engagement
- Personalize movie recommendations
- Increase movie viewing rate, recommend “long tail” items
Challenge
- Movie recommendations were made manually by the content managers department. Recommendation creation became a manual toil, regularly consuming significant human efforts.
- Some movies were not watched a single time.
Solution
Three types of movie recommender systems have been developed:
- The “Trending now” recommender offers users the most currently popular movies by the weighted average number of likes and view percentage.
- The “Movie-to-movie” similarity recommender offers the most similar movies to the movie that just has been watched by a user.
- Hybrid collaborative filtering recommender combining movie features, user features, and interaction data of users’ likes/dislikes and users’ movie watch percentages.
Impact
- Manual toil (content management by hand) was reduced by 60% and boiled down mainly to validate the automatically generated recommendations.
- Over one month of A/B testing of the integrated movie recommender system, the movie watch rate increased from 20% to 40%. Compared to the controlled deployment with manual recommendations, the average user screen time increased by 20%.
- The number of non-watched films decreased by 25%
RESEARCH /Molecular binding affinity prediction
Research
Reference publications:
-
Ensembling machine learning models to boost molecular affinity prediction
Ensembling machine learning models to boost molecular affinity prediction
-
High throughput screening with machine learning
High throughput screening with machine learning
-
Graph Neural Networks for Binding Affinity Prediction
Graph Neural Networks for Binding Affinity Prediction
Business Goals
- Cut expenses and duration of early drug discovery phases - target discovery and validation, lead identification and optimization.
- Extend virtual screening capabilities and accuracy by rejecting dubious molecular coordinates and transition to a more efficient parameterization of (bio)molecules.
Challenge
- Experimental methods of measuring binding affinity and inhibition constants are expensive in terms of required human efforts, time, and resources due to the tremendous number of chemical compounds.
- There are well-adopted methods of virtual screening, but they either lack generalization capabilities (AutoDock) or are extremely time and computational resources demanding (molecular dynamics)
Solution
- We propose a machine learning-based predictor for protein-ligand binding affinities.
- The pipeline unites two subsequent ensembles – classification and regression.
- Within the approach, the binding class and the binding strength can be assessed.
- We show that the use of diverse methods improves the prediction metrics.
Impact
- The binding affinity estimation MAE is ~1.5 kcal/mol of the Gibbs free energy. Taking into account the energy of water-water hydrogen bond O−H···:O (5.0 kcal/mol), we can conclude that it is a fairly accurate estimate.
- The inference time is 1 ms - not hours, days or weeks as is usually the case in molecular dynamics simulations.
WEBINAR AI for Pharmaceuticals and Biotechnology
Attending this conference will give you a comprehensive outlook on the key issues surrounding the latest AI technologies in pharma & biotech
https://www.eventbrite.com/e/ai-for-pharmaceuticals-and-biotechnology-tickets-352573736537?aff=ebdshpsearchautocomplete&keep_tld=1
About this event
The AI for Pharmaceuticals and Biotechnology 2022 conference will provide insight into the current state of the industries in the EU and US. In a multi-stakeholder setting, the event stimulates debate on the impact of AI technologies on the sustainability of pharma and biotech systems. Beyond a comprehensive outlook of market access policies, our speakers will outline recent critical advances in AI technology and their impact on the pharma and biotech industry.
Attending this conference will give you a comprehensive outlook on the key issues surrounding the latest AI technologies in pharma and biotech. This event will provide an essential platform for stakeholders to discuss and share best practices in furthering technology development.
The conference topics include but are not limited to the following:
- AI in Life Sciences - market overview and analysis
- AI in early clinical development and drug discovery
- Architecting auto-scalable domain data platforms
- Availability, integrity, and consistency of the data
- Modern multi-omics data analysis
- Pharmacogenomic drug profiling
- Overcoming aging with the assistance of AI
Reviews
the project
Custom Software Development for Medical Diagnostics Company
“The way blackthorn.ai stepped in and helped us revamp the model was excellent.”
the reviewer
the review
A Clutch analyst personally interviewed this client over the phone. Below is an edited transcript.
Introduce your business and what you do there.
I’m an executive at a medical diagnostics company. My division works on therapeutic research through small-molecule drug discovery.
What challenge were you trying to address with blackthorn.ai?
We needed support to build and train a machine-learning model. We started the project development but weren’t able to finish it.
What was the scope of their involvement?
blackthorn.ai has designed and developed custom AI software. At the beginning of our engagement, blackthorn.ai created a detailed statement of work (SoW), and any time they update it, they give us a summary of the aspects they’ll change.
We had started developing the machine-learning model ourselves but couldn’t develop autoencoders for a specific data type. We’ve worked in close collaboration with blackthorn.ai since there are many variables in the code. We have to be able to pivot and train models as we go.
Our project involves machine-learning algorithms, training models, autoencoders, and big data architecture. blackthorn.ai explains every possible pathway, lays down the options, and executes accordingly.
The machine-learning model blackthorn.ai has developed uses around 40–50 different data sources to find ways to develop new drugs without risking our patent. As we evolve with the project, the model will create new modules. Our goal is to accelerate our drug discovery timeline.
blackthorn.ai provides input from an outside point of view since they have staff with chemistry, biology, and mathematics backgrounds. blackthorn.ai has helped us create specific machine-learning models, and they’re currently working on data wrangling.
What is the team composition?
We started working with 2–3 teammates from blackthorn.ai, and the team composition has changed throughout our engagement. We currently work with around five teammates, including a manager, different software developers, and Alex (CEO). We’ve also worked with different specialists as needed. For instance, we needed mathematics specialists to develop one of the modules.
All the people I work with are in high roles, and they can delegate work to other specialists as needed.
How did you come to work with blackthorn.ai?
I found blackthorn.ai through a friend’s recommendation. We received other business proposals and liked the content and architecture of blackthorn.ai’s proposal.
How much have you invested with them?
We’ve spent around $250,000–$500,000 with blackthorn.ai. We plan to increase this budget since we just realized we’ll need to add more staff to our team.
What is the status of this engagement?
We started working with blackthorn.ai in July 2022, and our partnership is ongoing. We had an introductory meeting in June 2022, they sent us their SoW on July 2022, and we’ve been working together since then.
What evidence can you share that demonstrates the impact of the engagement?
I really like that blackthorn.ai creates detailed SoW and is good at giving updates. We’ve worked with other companies in the past, and when they pivot and change, sometimes the changes aren’t clear to them, even though we had previously discussed them verbally.
Before working with blackthorn.ai, we had developed our machine-learning model up until the point of clinical trials, but we had trouble getting past that. The way blackthorn.ai stepped in and helped us revamp the model was excellent. blackthorn.ai has developed in a couple of weeks what we tried to develop for about a month and a half.
Before working with blackthorn.ai, we tried to hire internal developers. However, blackthorn.ai’s staff background really stood out. They really helped us with the model’s autoencoders. I’m a chemist with a heavy programming background, and I don’t think we’d ever figured out how they architect these autoencoders. We’re actually thinking about publishing papers with them about this case.
Overall, our collaboration with blackthorn.ai has been great. They’ve developed detailed code and created great reports, which allow anybody from their team to continue the work at any given point.
How did blackthorn.ai perform from a project management standpoint?
blackthorn.ai exceeds our timeline expectations. Whenever they say they’ll deliver something by a certain date, they keep their word and deliver it even ahead of time on some occasions. I don’t need to be on top of them because blackthorn.ai works autonomously, which I appreciate.
We don’t have to meet with blackthorn.ai that often. They’re always descriptive in their emails about the work they’re planning to do. We have a quick check-up meeting with blackthorn.ai every other Tuesday, and if they run into a problem in the meantime, they email us so we can jump into a call. We also have a long call with the development team, usually about 1–2 hours long.
Moreover, blackthorn.ai’s recently sent an SoW for our extended engagement that includes the revision history, the list of data sets, the implementation roadmap, the responsibilities of each party, and an executive summary at the end. Every detail is pointed out, which is great for us since we’re a small team, and this way, we can manage our work more efficiently.
What did you find most impressive about them?
I’m most impressed with blackthorn.ai’s honesty. When we come up with an idea and blackthorn.ai thinks it won’t work, they tell us upfront and propose new ideas the next time we meet.
Are there any areas they could improve?
The only constructive feedback I can give blackthorn.ai is that they hesitated to suggest ideas at the beginning of our engagement. However, that has changed throughout our partnership, and blackthorn.ai has been great since. I told Alex this was a collaborative relationship, and their proposed solutions have worked most of the time.
Do you have any advice for potential customers?
I suggest approaching blackthorn.ai with a clear picture of what you want to accomplish. You don’t need to know all the details, but be clear about your project goal. I also recommend asking blackthorn.ai for their input during the project and engaging with them. They might not be the right fit for every customer, but when they send you their initial SoW, you’ll know if they’re right for you.
the project
Customized Solution for OTT Company
"All goals were achieved."
the reviewer
the review
The client submitted this review online.
Please describe your company and your position there.
I'm the CTO of an over the top (OTT) company. We provide TV and VOD service in 27 countries.
For what projects/services did your company hire blackthorn.ai, and what were your goals?
- Modernize the company’s current approach of analytical data collection, storage, and preprocessing as well as interfacing with existing users.
- Achieve competitive advantage in Total Cost of Ownership (TCO) by moving to the Cloud-based SaaS model.
- Offload the DBA team by automating the BI as well as providing access to the required data and analytical tools to a wide range of end-users so they can do the analytics on their own.
How did you select this vendor and what were the deciding factors?
The team has a lot of certificates including Amazon and Google cloud technologies, and works with big data.
The company's approach at the research stage made it clear that the project would be implemented with high quality and on time.
Describe the scope of work in detail, including the project steps, key deliverables, and technologies used.
- Total Cost of Ownership (TCO) was carefully considered by comparing the performance and costs of alternative services offered by a number of cloud providers. The most advantageous combination of cost, quality, and sustainability was chosen to meet customer requirements.
- Business requests were collected by interviewing numerous stakeholders. The requests served as a starting point for the BI dashboards, the DWH schema, and data pipelines design.
- The multidimensional analytical Data Warehouse (or OLAP cube with facts, dimensions, etc.) was built and extended piece-by-piece revealing one business request after another. The cloud warehouse consolidated large amounts of data from disparate on-premises sources.
- Convenient BI dashboards were designed and implemented to meet collected earlier business requests and their possible future variability. The end business users with no SQL skills could perform requests using straightforward BI dashboards, without DBA team involvement.
- Auto-scalable ETL workflows were designed and implemented as the business requests and DWH schema had been established. The pipelines extract, transform, and load data from the source databases (on-premises) into the target DWH (on the cloud).
- An incremental index was introduced to large source tables to upload only updates, not complete tables.
- Backward historical consistency was achieved by introducing meta-tables and unique timestamps that indicated the time range when a source row or table was actual.
How many people from the vendor's team worked with you, and what were their positions?
The team consisted of:
- Solutions Architect
- Responsible for the project success
- Business requirements collection and analysis - System analysis and design
- System requirements specification
- Project management
- Team technical leadership
- Sr. Data Architect
- Technical leadership of the project
- DWH schema design and implementation - ETL schema and workloads design
- BA/BI reports design
- Code review
- Data Engineer
- Executes the project plan
- Implements DWH as designed by Data Architect
- Implements ETL workloads as designed by Data Architect
- DevOps
- Deploys cloud resources
- Develops scripts for automated deployment of the environments (dev/prod)
- Manages user and service accounts permissions
- Business Intelligence Engineer
- Implements BI dashboards and scheduled emailing as designed by Data Architect.
Can you share any measurable outcomes of the project or general feedback about the deliverables?
All goals were achieved.
We have accelerated the process of generating reports day by day, previously the process took an average of a week.
The analytics team received a flexible reporting solution.
We have optimized the workload of developers and freed up about 300 hours per month per team thanks to the new solution.
Also, thanks to the cloud solution, we did not need additional Devops to maintain the infrastructure.
Describe their project management style, including communication tools and timeliness.
Blackthorn’s team uses an open management style, they are open to discuss any project question.
They provide video recordings for every meet, make a summary of the meeting with the step plan. They always use a tracking system for the tasks.
Blackthorn’s team communicates via messengers, email and are available 24/7 for the dialog and solving questions!
What did you find most impressive or unique about this company?
Exceptional communication, great work, efficiently and extremely knowledgeable!
Are there any areas for improvement or something they could have done differently?
I wish the team to grow and open new directions!
the project
AI Backend System Development for Skincare Application
“They stand out for their competitive advantages and expertise.”
the reviewer
the review
A Clutch analyst personally interviewed this client over the phone. Below is an edited transcript.
Introduce your business and what you do there.
I’m the CEO of a skincare company; we have some brands. Currently, we’re developing a personalized skin treatment solution. We have a connected device, an application, and an AI system to detect skin conditions and recommend products.
What challenge were you trying to address with blackthorn.ai?
We wanted to build a reliable skin condition classifier. The system would take an image, recognize the person's skin condition, and convert it into a diagnostic. Our challenge was making a personalized product that could be used on a mobile device at home.
What was the scope of their involvement?
Initially, blackthorn.ai created the architecture for the AI system. They built models of skin classification from scratch. Several models were under investigation, and we selected the best one in the end. Then, they tested, implemented, and deployed the model that’s now used in our mobile application. blackthorn.ai didn’t work on software development; they did the backend on the AI and machine learning side. We provided briefs, and their team worked with our team of dermatologists to align with the actual skin conditions.
The system is connected to a mobile app. Users have a simple interface; they can go in, answer questions, and scan their skin via their phones. The system takes an image, which is further analyzed by the AI system to provide the results.
What is the team composition?
We’ve worked with about 4–5 people, including the main architect, a computer vision specialist, and a DevOps person.
How did you come to work with blackthorn.ai?
Another company recommended them. Before we started working, we got acquainted with their chief solution architect, who is also the company’s CEO.
How much have you invested with them?
We’ve spent about $100,000.
What is the status of this engagement?
The official scope of work started in September 2021. We ended the first part of the project, but we're now investigating the next phase, so the partnership is still ongoing. They’re currently offering a few hours per month of support.
What evidence can you share that demonstrates the impact of the engagement?
We're very satisfied with their work in general. We ran blackthorn.ai’s model against an open-source code that was available in the market by another large corporate and found that their model’s results were 18% better than the other one.
How did blackthorn.ai perform from a project management standpoint?
Project management is excellent. Everything has been delivered on time and well documented. Communication is good. During development, we had weekly updates. If there were any sort of delay, it was always shared — although we didn’t really have delays. Additionally, we used Jira to track all the tasks.
What did you find most impressive about them?
They stand out for their competitive advantages and expertise. Moreover, they deliver what they promise and don’t promise what they can’t deliver.
Are there any areas they could improve?
No, there’s nothing to change. They should just move in the same manner.
Any advice for potential customers?
Be very precise at communicating the goals you want to achieve with your products; the clearer you are, the better.
blackthorn.ai has helped the client resolve a development problem they couldn’t solve. They provide detailed SoWs and have exceeded the client’s timeline expectations. blackthorn.ai stays on budget and works autonomously. They have strong and honest communication through email.