Adept ID

Changing jobs between industries is incredibly hard, particularly for the 80m working Americans who haven’t gone to college. Meanwhile, employers in growing sectors like Healthcare and Renewable Energy can’t hire fast enough from their traditional sources of applicants.

Fortunately, a person’s past work may have given them many of the skills they need to succeed in a seemingly different role. If employers could find these people, they would hire faster and more inclusively.

AdeptID uses machine learning on underlying skills data to identify non-obvious, high-likelihood transitions, specifically for roles that don’t require college degrees.

As employers and training providers use our analysis, they reduce their time to hire and make their processes more inclusive of non-obvious talent. With more adoption, AdeptID ingests more outcomes data that further improves our insights, which we can share with job seekers and training providers in the form of life-changing recommendations.

An estimated 35 million American workers without college degrees are either unemployed or working in industries that are in structural decline. At the same time, employers in high-growth sectors such as allied health and renewable energy cannot find enough qualified applicants to meet their needs.

This disconnect between demand and supply is evident in training and hiring processes. Vocational training programs have very  low job placement rates (typically 35%). Hiring managers are spending disproportionate time screening candidates they know will not be qualified, further highlighting the inadequacy of the status quo.

We believe that the existing infrastructure for job transitions is so ineffective in large part because it is one-size-fits-all. We see the challenge of reskilling as a matching problem - very similar to ones that data science has been able to solve in other contexts. All individuals have a latent set of skills, capabilities, and aspirations that make them more or less likely to succeed in jobs, even jobs that are superficially different. If we can properly capture the real demand signals from employers, we can recommend high-impact, non-obvious pathways and dramatically improve the efficacy of the reskilling space.

AdeptID is building a recommendation engine that connects individuals to in-demand jobs and the training to get them there. We use machine learning to find latent, transferable skills individuals have developed in past occupations that predict success. We then identify non-obvious, high-impact job transitions, which we can recommend to employers, training providers, and the individuals themselves.

The recommendation engine is the novel core to our solution, but we put it to use in different, complementary ways:

• Employers use our technology to receive applicant scoring and analysis that help them recognize new candidates suited for their in-demand roles. They also use our technology to identify workers on their staff who may be eligible for advancement.

• For Training Providers, our recommendations inform recruitment, learner pathing, and placement.

• Finally, individual job seekers receive insights into their own latent transferable skills surfaced by our models thanks to the data collected above. They can also receive recommendations and connections to optimal training and employers.

As employers and training providers return data on outcomes, the models dynamically improve and further remove friction from the job matching process.

Our special sauce is back-end infrastructure, but we deliver our insights via API, data files, and interactive dashboards.

AdeptID’s mission is to make job transitions easier for the 80 million working Americans without college degrees. This group has been particularly vulnerable to displacement. Longer working lives and increased rate of technological change are a recipe for personal and societal disaster if we don't make it easier to recognize transferable skills.

We think the best way for us to start to serve this population is by making it easier for employers to hire them. That’s why our initial product is focused on the employer pain point of skills gaps for growing middle-skilled jobs. We need to understand the real demand signals coming from these employers in order to provide hope of advancement for the workforce itself.

Once we capture these demand signals, we can empower individuals with recommendations to the training programs and employment opportunities best suited for them.

We also believe our use of collaborative filtering - a machine learning technique that has fueled successful personalization at companies like Netflix and StitchFix - can help make job transitions more personal and successful for this segment.

We have also intentionally chosen to work with Healthcare and Renewable Energy - industries that are economically and socially sustainable.

Our technology matches current and future employers' needs to talent in a way that is sensitive to the abilities and potential of each individual.

Because our models will implicitly assess the impact of training programs and providers, as well as jobs and employers, we will also help learners make decisions about pathways with the highest ROI. In addition, our ability to quantify the impact of underlying skills and attributes means that we will be able to empirically surface issues of bias for remediation.

We have 2 full-time team members (Fernando Rodriguez-Villa and Brian DeAngelis) and 1 part-time member (Subit Chakrabarti). Over the next 12 months, we plan to bring 2-3 more data scientists and engineers to build out our solution.

The problems we are trying to solve are too massive and complex for us not to require a diverse team that is inclusive of different perspectives.

As founders, we have been fortunate in our academic and professional associations. We are, nevertheless, products of families from diverse educational and ethnic backgrounds. As children of immigrants, we have experienced firsthand the potential of a dynamic and inclusive society. Our recognition that innovation is required to preserve that society led to the founding of AdeptID.

We believe that everyone is adept. We all have skills that make us more versatile than past job titles might suggest. We are following some tactical ways to make sure our company lives up to that creed: skills-based hiring, B-incorporation, participation in minority talent boards, a generous employee equity pool. Active partnership with New Profit and the Morgridge Family will ensure we keep our best feet forward.

The concept of matching jobs on the basis of underlying skills is not novel. Labor market data providers such as Emsi and Burning Glass provide “skill similarities” between jobs. These insights are only descriptive of differences - they make no effort to predict an outcome (of successful job transition). As a result, they cannot make claims of accuracy or efficacy and are less helpful to the user - whether that’s an employer, training provider, or job seeker.

To our knowledge (and we’ve been searching pretty hard), no one in the market has collected the real hiring data from employers and trained models on those outcomes - certainly not for the middle-skilled labor segment, which we believe is distinct enough to warrant its own solution.

AdeptID’s approach of simultaneously a) serving the needs of and b) collecting data from employers allows us to differentiate ourselves on the basis of both business model and technology. Our focus on middle-skilled roles in the specific sectors of Healthcare and Renewable Energy further makes us stand out.

Customers will need to believe that our solution has an ROI on top of the status quo - in the case of employers, that includes an Applicant Tracking System, a Talent Acquisition team performing manual screening, and occasionally an outsourced sourcing team. The good news is that we can demonstrate ROI empirically - our systems can draw straight lines between model accuracy and customer-defined KPIs.

Our technology predicts the probability of successfully transitioning to a new occupation based on existing, transferable skills. To build this solution, we’re leveraging machine learning techniques that are used elsewhere to run powerful recommendation engines (e.g. Netflix, Stitchfix). Our technical approach involves two key activities - feature engineering and model training:

1. Feature Engineering We capture the similarities and differences between occupations by looking at their underlying skills. To do this, we’ve ingested publically available skill attributes from the Bureau of Labor Statistics (BLS) as well as several proprietary providers to create a large table describing ~1,000 occupations using ~10,000 distinct skills. We then use Singular Value Decomposition (SVD) to construct a smaller set of latent skill features that describe the overall similarity between occupations. SVD uses the correlation structure of the original skill attributes to surface latent variables that capture the maximum amount of variance between occupations in the fewest number of features. We then use an individual’s work history to describe each individual in the space of latent skills.

2. Model Training Using these latent variables, we train classification models to predict outcomes of interest (successful hire, successful retention, promotion etc.) using historic employer outcomes data. These trained models then allow us to recommend high-potential applicants with non-traditional backgrounds to employers or training providers. With out-of-sample testing, we can then statistically demonstrate the efficacy of these models.

The machine learning approach described above leverages techniques that have been widely successful in building production grade recommendation engines. Two well documented examples are the Netflix Prize and Stitch Fix’s Style Engine. Though they have yet to be deployed in a skills context, there is little technical risk in the underlying techniques. Because we are using standard, linear techniques (SVD, Logistic regression, SVMs), our trained models can be directly inspected - removing the “black-box” nature of many machine learning approaches as well as allowing us to better demonstrate the absence of bias in our predictions.

When we’ve run our models on real hiring data we’ve received, we have been able to demonstrate some encouraging predictive power. While our results are constrained by the quantity of data we have to date, our models are already predicting hiring outcomes at 75-80% accuracy - just on the basis of skills fit. In a subset of cases where we have high confidence, we are seeing >90% accuracy - this will have dramatic implications for those prioritizing resources for hiring and training.

As we get more data, we can use more predictors, which should lead not only to increased accuracy but also more nuance in our recommendations.

Our solution drives impact by promoting Transparency and Personalization.

The “matching process” between talent and demand is opaque and homogenous. Identifying and recommending individuals on the basis of latent attributes leads to more transparency for all parties.

Because our insights work at an individual level, we can make sure that the path for each person is unique to them. No more homogenous approach to job search, training, or hiring. 

Outcomes

In the short-term, we want employers to recognize that previously overlooked pools of talent can fill their in-demand roles. This will happen as they observe results (successful hires) recommended by our models. Our models can also shift behavior by making it faster for employers to find talent from (speed to hire is often their primary motivation).

Adoption will lead, in the medium-term, to new methods of hiring oriented more around desired skills and roles (“skills-based hiring”). These practices will weaken and in some cases obviate some of the barriers (degrees) that have made transitions so painful and inefficient.

In the long-term, our solution can redefine the way talent is recognized in every segment of the labor force. By recognizing the adeptness and potential of every individual, we want to become the talent agent and cheerleader for workers everywhere. We believe this labor empowerment will promote an inclusive society and politics that reflect optimism and potential.

Inputs

Our inputs include our strategy to partner with employers to understand demand, our activities to collect outcomes data, and our interventions to provide recommendations and scoring for transitions at the employer and training provider level.

We need to invest our time as founders into understanding these partners so that our technology fits in their workflow and can drive behavior change without causing undue burden. 

Outputs

Our solution directly impacts individuals hired by our employer partners and those trained and placed by our training provider partners.

Beyond being scalable, our solution benefits from scale. The more partners, the more data, the smarter our impact is - leading to not just a higher volume of placements, but a higher rate of them.

Over the next 12 months, we will grow our list of employer and training provider partners, who will utilize our models on a daily basis to find, train, and hire promising middle-skilled workers.

We believe our models can influence the training and hiring of over 10,000 people in the next year. We want to demonstrate additionality in >10% of these cases.

In that time period, we want to have collected enough real observation data from customers so that our recommendations are based on our own novel data asset (rather than conventional sources such as the BLS or job postings aggregation sites).

By mid-2020, we’ll have established an Advisory Board with representative experts from the fields of Workforce Development and Talent Acquisition. These will ensure that our solution is built and deployed for maximum impact.

Over the next five years, we want to scale our Recruiting and Placement Analytics to play a role in the training and/or hiring of over 1 million workers (~35% of the annual job openings in our targets areas of allied health, renewable energy, and skilled trades) and demonstrate additionality for >20% of these cases.

As we reach key business and development milestones, we will also offer more “full-stack” Placement and Admission Services across industries. Our platform will allow us to serve as a Frictionless Talent Broker and occasional Capital Provider for the 80m working Americans without college degrees. Within five years, we want our models to have helped obtain credit for >100,000 individuals.

A great deal of inertia has led the labor market to perform the way it does. Businesses haven’t been accustomed to investing in technology to solve even well-known hiring problems. Even if we do a great job finding our way into their workflow, a fair amount of behavior change is necessary. Furthermore, while certification and credentials have important signalling value, it is likely that these also present barriers that prevent transitions that would otherwise be advantageous. 

We also see two major technical barriers:

• Data accessibility and data quality due to the fragmented nature of the market: A major challenge in the re-skilling space is the fragmented nature of the market (of training providers and employers and occupations). 

• Inherent predictability of the problem: Our current approach uses past decisions and behavior to understand individuals’ likelihood of success. We have strong reason to believe that these outcomes are influenced by legible predictors (e.g. attributes of the candidate, macroeconomic factors, attributes of the company etc.), so long as we can find them and incorporate them. However, it’s possible that randomness plays a large role in this historic outcomes data. The greater the degree of randomness in these historical decisions, the greater the unpredictability of outcomes.

Business and GTM

A silver lining of the pandemic, economic crisis, and social justice protests of 2020 is that inertia has been stopped: Many businesses are looking for an approach to staffing that improves on the status quo - this is the moment for bold solutions. Healthcare and renewable energy are in secular growth, and are feeling the pain of hiring enough that they are uniquely willing to invest in our solution. Over time (we believe in the 12-24 month timeframe), our results will be able to speak for themselves and we will move from the early adopters to the early majority (and other industry segments).

Data Accessibility and Quality

Our approach represents multiple occupations, training providers, and employers in a single general model. In our first pilots, we can determine if a model of this nature is predictive enough to improve employer KPIs. Because sparsity is expected in collaborative filtering, we will not need all data from all providers to be fully populated. This collaborative filtering approach allows us to avoid most data quality problems and demonstrate performance.

Inherent Predictability 

We believe that if we can obtain and ingest the following types of data, we will be able to demonstrate differentiated performance:

• latent skills representation for occupations,

• representation space for training providers,

• representation space for employers,

• psychometric characteristics,

• geography, demographic information, and

• legible interventions (i.e. actions taken to improve outcomes).

Our models are built using historical outcomes data from partner institutions. It is currently difficult to gather longitudinal data that tracks individuals over many years especially in contexts where they’ve switched employers. Data of this nature is incredibly valuable because it allows us to measure longer term outcomes and implications for particular career transitions.

To address this limitation, we’ve partnered with nonprofits like YearUp who have made investments over decades to track the outcomes of participants in their program. As we deepen our engagement with employers and build tooling for individuals, we will, over time, be able to collect this longitudinal information on our users.

Our team has the combination of entrepreneurial and technical expertise required to implement our proposed solution. As business and technical leaders in machine learning ventures, we have worked to match the awesome potential of data-driven software with the practical needs of users.

Fernando’s work with large financial institutions at JP Morgan gave him insight into how novel approaches to risk management could create value for underrepresented segments, and how large partnerships are forged between organizations. He signed Knewton’s first partnerships with digital education providers in India, South Africa, and Southern Europe. As the GM and business lead at TellusLabs, he grew the business from $0 to >$2m in revenue within 18 months by working closely with business customers to scope the offering to their needs.

As a data scientist, Brian has experience in both industry and academia building production-grade models from diverse and unwieldy data sets. His models have underwritten novel approaches to quantifying soil carbon credits at Indigo Ag. Brian has a PhD in Computational Neuroscience from Yale where he developed novel computer vision methods for markerless tracking of insects.

We have learned that successful adoption of data-driven applications requires great stakeholder management as well as great analysis. AdeptID’s solution will require plenty of both, and we’re eager for the challenge.

We are working with Boston Medical Center, a safety net hospital here in Boston that hires over 300 middle-skilled roles a year. They have provided extensive past hiring data for 4,500 job openings and over 100,000 applications, which has allowed us to better understand the hiring dynamics for in-demand healthcare workers. We are helping them identify new pools of talent for pharmacy technicians, medical assistants, and CNAs.

We have also signed a partnership agreement with Emsi that gives us access to their best-in-class labor market data and provides for cross-marketing of our data and analytics services.

We have also signed partnership agreements with YearUp and Jewish Vocational Services, which have shared thousands of anonymised job placement outcomes, which have been pivotal in allowing us to train our initial models.

We have other agreements in place or in negotiation with partners that we are unable to name in this public submission.

We have three beneficiary segments, to whom we offer complementary value propositions:

• Employers: Access to non-obvious pools of talent (particularly for hard-to-fill roles); Faster, more efficient hiring of middle-skilled workers; Insights into optimal hiring practices

• Training providers: Improved efficiency and targeting of learner recruitment, pathing, and job placement.

• Job-seekers without college degrees: Matching to new opportunities on the basis of latent, transferable attributes.

We are initially monetizing via the Employers. In the near-term, our business model focuses on selling Recruiting Analytics Services to employers in the healthcare and renewable energy industries. This is where our value creation will be easiest to monetize (and who has the capacity to pay). In some cases (like in our early work with a renewable energy company), those recruitment analytics are being implemented upstream with the training providers that train talent for the employer.

Right now, we are charging from $2,500 - $20,000 a month for our analytics services. As we demonstrate the relationship between our models accuracy and customer KPI’s, we will move to a per scored user pricing model (we think we will charge $100-$200 per placement, which would lead to >$100k annual customer values).

As we reach key business and technology milestones, we will begin to offer Placement and Admission Services to businesses in these and other sectors. Ultimately, our platform will allow us to serve as a Talent Agent and occasional Capital Provider for the 80m working Americans without college degrees.

We will fund our operations through the sale of products and services (following the business model laid out above). Our plan is to have monthly revenue receipts that cover costs by year-end 2021.

We will periodically seek external investors, but for growth purposes only once we have identified clear opportunities that require capital that we cannot immediately generate from operating activities.

Any funding we can secure through mission-aligned, impact-oriented sources, such at MIT Solve, New Profit, and the Morgridge Family will allow us to accelerate progress towards our goals.

We have not raised external funds for our solution, though we do have investors in-network who have expressed interest in participating once we conduct a formal round of funding.

In the interest of discretion, we are happy to share commercial results and progress only in a private context.

Based on recent technical and commercial results, we are evaluating options for initial funding. From past experience, we know the potential of traditional high-growth funding sources to accelerate a business. We also recognize that these funding relationships are only successful when the investor and the operator are aligned on mission and objectives.

We believe that investors have historically compartmentalized their approach between projects that are socially important and projects that are potentially lucrative - a practice that is finally evolving. We are looking for investors that reject that distinction and will be true partners in helping us build an impactful, valuable enterprise.

We estimate expenses between $1m and $1.5m in 2021, depending on the amount of commercial traction and timing of external funding events.

The majority of our expenses will be wages for staff, with minor budget requirements for ingredient data purchases, software subscriptions, and travel.

When we were first shown the challenge prompt, we felt like we were reading a rearticulation of the same thinking that led to the founding of AdeptID. From what we can tell, we are incredibly well aligned in acknowledging the gravity and pervasiveness of the problem, but also in being optimistic that technology can rewrite the script and promote inclusive and personalized employment experiences for a previously overlooked segment of the population.

Our social-impact venture is at an exciting yet precarious moment. Our models are starting to show some really exciting results and we have the attention of a few important partner organizations. However, there is a ton of work in front of us to better understand the workflows of our partners and to make sure that we are ultimately meeting stakeholders, particularly job seekers, where they are. The validation pilots described seem like a perfect proving ground for our technology and a chance for us to really sync with workforce needs.

Despite our passion for the problem, we are functional - and not subject matter - experts. Participation in this challenge presents an opportunity for us to work with organizations with deep expertise that complements ours. We hope to learn from the experience of New Profit, the Morgridge Foundation, and MIT Solve in this space.

We are building fundamental technology that we believe will have broad application throughout the workforce development space. As any early venture, we must focus our efforts in a way that allows us to address the massive problem we're solving while ensuring the progress/health of the company. We need partners who will help us focus on the right near-term objectives to achieve our long-term goals. In our case, that particularly involves getting people "at the table" who understand the workforce development and staffing spaces.

Given our small team, any assistance on increasing our exposure would improve the odds of us finding the right partners to scale rapidly.

Successful implementation of our solution depends on healthy partnerships with a range of mission-aligned institutions.

We would love to partner more actively with workforce boards and other aggregators of supply so that we can share our recommendations more directly to those seeking work. Our early strategy has dictated that we focus on understanding demand signals from employers, but we will be ready soon to offer insights directly to people seeking work. This is most likely to succeed if we find the right partners to grow our footprint (hopefully from the large networks of the challenge sponsors and other participants). 

We hope to expand our current work with healthcare employers. While we have been successful in getting to know the workforce development priorities of Massachusetts-based institutions like Boston Medical Center and UMass Memorial, we would like to build relationships with large systems across the country, particularly in non-coastal states.

Similarly, we want to build on what we’ve learned with our current partners about pathways to employment as wind turbine technicians and other renewable energy roles, especially in regions of the country where economic rehabilitation is critically needed. Partnering with other large employers and training institutions focused on these problems would help us grow our training data set and get our insights into the hands of more important players.