Denis Ignatovich, Co-founder and Co-CEO of Imandra, has over a decade of expertise in buying and selling, danger administration, quantitative modeling, and sophisticated buying and selling system design. Earlier than founding Imandra, he led the central danger buying and selling desk at Deutsche Financial institution London, the place he acknowledged the essential function AI can play within the monetary sector. His insights throughout this time helped form Imandra’s suite of economic merchandise. Denis’ contributions to computational logic for monetary buying and selling platforms embrace a number of patents. He holds an MSc in Finance from the London College of Economics and levels in Pc Science and Finance from UT Austin.
Imandra is an AI-powered reasoning engine that makes use of neurosymbolic AI to automate the verification and optimization of complicated algorithms, significantly in monetary buying and selling and software program programs. By combining symbolic reasoning with machine studying, it enhances security, compliance, and effectivity, serving to establishments cut back danger and enhance transparency in AI-driven decision-making.
What impressed you and Dr. Grant Passmore to co-found Imandra, and the way did your backgrounds affect the imaginative and prescient for the corporate?
After faculty I went into quantitative buying and selling and ended up in London. Grant did his PhD in Edinburgh after which moved to Cambridge to work on purposes of automated logical reasoning for evaluation of security of autopilot programs (complicated algorithms which contain nonlinear computation). In my work, I additionally handled complicated algorithms with numerous nonlinear computation and we realized that there’s a deep connection between these two fields. The way in which that finance was creating such algorithms was actually problematic (as highlighted by many information tales coping with “algo glitches”), so we got down to change that by empowering engineers in finance with automated logical instruments to deliver rigorous scientific strategies to the software program design and growth. Nevertheless, what we ended up creating is industry-agnostic.
Are you able to clarify what neurosymbolic AI is and the way it differs from conventional AI approaches?
The sphere of AI has (very roughly!) two areas: statistical (which incorporates LLMs) and symbolic (aka automated reasoning). Statistical AI is unbelievable at figuring out patterns and doing translation utilizing data it realized from the information it was skilled on. However, it’s unhealthy at logical reasoning. The symbolic AI is sort of the precise reverse – it forces you to be very exact (mathematically) with what you’re making an attempt to do, however it could possibly use logic to motive in a means that’s (1) logically constant and (2) doesn’t require information for coaching. The strategies combining these two areas of AI are known as “neurosymbolic”. One well-known software of this method is the AlphaFold venture from DeepMind which lately gained the Nobel prize.
What do you assume units Imandra aside in main the neurosymbolic AI revolution?
There are numerous unbelievable symbolic reasoners on the market (most in academia) that focus on particular niches (e.g. protein folding), however Imandra empowers builders to research algorithms with unprecedented automation which has a lot larger purposes and larger goal audiences than these instruments.
How does Imandra’s automated reasoning eradicate widespread AI challenges, comparable to hallucinations, and enhance belief in AI programs?
With our method, LLMs are used to translate people’ requests into formal logic which is then analyzed by the reasoning engine with full logical audit path. Whereas translation errors could happen when utilizing the LLM, the person is supplied with a logical clarification of how the inputs have been translated and the logical audits could also be verified by third celebration open supply software program. Our final purpose is to deliver actionable transparency, the place the AI programs can clarify their reasoning in a means that’s independently logically verifiable.
Imandra is utilized by Goldman Sachs and DARPA, amongst others. Are you able to share a real-world instance of how your expertise solved a posh drawback?
An ideal public instance of the actual world influence of Imandra is highlighted in our UBS Way forward for Finance competitors 1st place win (the main points with Imandra code is on our web site). Whereas making a case research for UBS that encoded a regulatory doc that they submitted to the SEC, Imandra recognized a elementary and refined flaw within the algorithm description. The flaw stemmed from refined logical circumstances that must be met to rank orders inside an order e book – one thing that might be unattainable for people to detect “by hand”. The financial institution awarded us 1st place (out of greater than 620 firms globally).
How has your expertise at Deutsche Financial institution formed Imandra’s purposes in monetary programs, and what’s probably the most impactful use case you’ve got seen to this point?
At Deutsche Financial institution we handled quite a lot of very complicated code that made automated buying and selling choices based mostly on numerous ML inputs, danger indicators, and many others. As any financial institution, we additionally needed to abide by quite a few laws. What Grant and I noticed was that this, on a mathematical stage, was similar to the analysis he was doing for autopilot security.
Past finance, which industries do you see as having the best potential to learn from neurosymbolic AI?
We’ve seen AlphaFold get the Nobel prize, so let’s positively rely that one… In the end, most purposes of AI will vastly profit by use of symbolic strategies, however particularly, we’re engaged on the next brokers that we are going to launch quickly: code evaluation (translating supply code into mathematical fashions), creating rigorous fashions from English-prose specs, reasoning about SysML fashions (language used to explain programs in safety-critical industries) and enterprise course of automation.
Imandra’s area decomposition is a novel function. Are you able to clarify the way it works and its significance in fixing complicated issues?
A query that each engineer thinks about when writing software program is “what the sting circumstances?”. When their job is QA and they should write unit take a look at circumstances or they’re writing code and interested by whether or not they’ve appropriately applied the necessities. Imandra brings scientific rigor to reply this query – it treats the code as a mathematical mannequin and symbolically analyzes all of its edge circumstances (whereas producing a proof concerning the completeness of protection). This function relies on a mathematical method known as ‘Cylindrical Algebraic Decomposition’, which we’ve “lifted” to algorithms at giant. It has saved numerous hours for our prospects in finance and uncovered essential errors. Now we’re bringing this function to engineers in all places.
How does Imandra combine with giant language fashions, and what new capabilities does this unlock for generative AI?
LLMs and Imandra work collectively to formalize human enter (whether or not it’s supply code, English prose, and many others), motive about it after which return the output in a means that’s simple to grasp. We use agentic frameworks (e.g. Langgraph) to orchestrate this work and ship the expertise as an agent that our prospects can use straight, or combine into their purposes or brokers. This symbiotic workflow addresses lots of the challenges of utilizing LLM-only AI instruments and extends their software past beforehand seen coaching information.
What’s your long-term imaginative and prescient for Imandra, and the way do you see it reworking AI purposes throughout industries?
We predict neurosymbolic strategies would be the basis that paves the way in which for us to comprehend the promise of AI. Symbolic strategies are the lacking ingredient for many of the industrial purposes of AI and we’re excited to be on the forefront of this subsequent chapter of AI.
Thanks for the nice interview, readers who want to be taught extra ought to go to Imandra.