Products are shipped once, and maintained forever. Problems naturally arise as the system interacts with its environment from unforeseen bugs to malicious third-party attackers. It's the job of the company to ingest the exhaust generated from these systems, and to solve issues as they appear — ideally in a low cost way, as quickly as possible.
Complications emerge rapidly because problems are structurally hard to find and, especially as products attract more users, occur in higher volumes. Oftentimes problems are adversarially hidden by a fraudster or hacker continuously creating new strategies to extract information or money.
Today, companies deploy a myriad of point solutions home-grown or purchased through a vendor to maintain their products. A typical solution might include buying "signals" from four startups and a fleet of human analysts triaging those signals. Another pattern is a fleet of human data scientists building detection algorithms and passing those signals to human analysts. Perhaps there are some auto-resolution workflows, depending on the alert type. Maintaining products today is costly, slow, and bottlenecked by human attentiveness.
The Underlying Data Problem
Companies are in an eternal war protecting their products. Customers light up call centers with new complaints ranging from a feature breaking to an account exploit and companies, caught on their back foot, rush to solve their problems. Incidentally, the company could have known about their problem from the start. The string of fraudulent transactions, the abnormal drift in the logs. It's all a data problem.
There's been a herculean effort to use internal data to solve these problems. In fact, the majority of the headcount at scaled up companies works on some part of this process. But the catch is any type of deterministic system, even one that exploits the company's internal data, is destined to only work for some period of time. The real world is not just unpredictable, it's adversarial. Problems continue to evolve as the company does — and in fact, the day a company's problems stop evolving, is the day the company has its most important problem.
How do you build an evolving system to automate detection? The system must automate the processing of data end-to-end from feature engineering to analysis. And the system must augment its strategy for each module of the data processing pipeline as it learns which strategies succeeded or failed in the real world.
Vertical Integration
Conway is vertically integrated. By automating detection, investigation, and decisioning in a single platform, Conway creates a feedback loop. Over time, the system gets better at identifying the signature of an attacker and can even adapt as the company's interface changes or the hacker uses an alternate strategy. There is no re-training period and no labor-intensive label generation, unlike traditional ML models that you might find in the biggest banks, insurance, telecom, and technology companies.
The Feedback Loop is the Primary Product
In addition to enabling self-adaptation, vertical integration allows for novel detection schemes. In traditional paradigms, certain algorithms might have a high false positive rate and get removed due to the operational burden — even though one out of 100 alerts was highly relevant and could not be caught with another method. By automating investigation, the cost of false positives is nearly zero, including both latency and cost.
Build Resilient Systems
We have many hard problems to solve — from understanding high dimensional, multi-modal data on the fly to learning iteratively from human and environmental feedback. Consider joining Conway if you're excited to work on them. Please reach out to careers@conway.ai for next steps.