Our Expertise
instrumentation
While the world fusses over models, statistics, predictions and metrics - journeys in data science/ AI often begins not with the model, but with the collection of data at a granular level that can make a good model. This task involves designing the sensor (e.g., data product) which detects the right data, and estimation of how sensed values become measures. Sometimes when the model does not improve in accuracy, better instrumentation is the answer.
data democratization
The more people interact with AI and the data that its derived from - the more people educate themselves about AI, uncover hidden biases, can tell stories powered by data insights and can suggest better analytics and visualizations. Data democratization involves building data lakes, creating powerful intelligent and interactive dashboards, views and customizable tools that let people play with the data.
ML Technical debts
Complex prediction systems come with their own risk factors that must be serviced. Technical debts in ML software are very different than traditional software. Thus, for all our understand and tracking non-ML software debts, we are only in the beginning of handling technical debts in ML. Issues we check for include fixing unstable data dependencies, catching for eroding abstraction boundaries, feedback loops, ML anti-patterns and upstream vs. downstream control etc.
epochs & data residue
As the model interacts with real world, it receives reinforcement and gets upgraded. But whats also left behind is frozen versions of the input data - what we call “data residue” - sequences of input/ output that engendered the model. Such data residue is critical to purifying model evolution, accuracy growth over epochs and monitoring positive and negative transfer of information across versions of the ML software. Its also critical to reproducibility and re-spawns.
model tuning
Just as humans improve their mental models over time, similarly machine learning models upgrade with more data, negative examples and adversarial learning. But how much time until a significant accuracy bump? And how much diversity of data is required? And what sort of hyper-parameter optimization is necessary for such improvements? We understand the importance of tuning, upgrading and continuous learning ML models - and design accordingly from the start.
explainability & rationale generation
Powerful machine learning models often trade-off accuracy with explainability. Complexity is no misnomer, and complex algorithms can’t always explain themselves. We build interrogation APIs that talk to the model to map how it will behave, and depending on the industry - the importance of providing explainable solutions that might be called upon by necessary authorities.