Overstone Spotlight
Siva Karanam: Senior Software Engineer
Could you tell us about your professional journey and what drew you to the field of data at Overstone? How has your experience shaped the way you approach data and analytics today?
Before Overstone, I worked at Rakuten, a Japanese technology group, and at Myntra, one of India’s largest fashion e-commerce platforms. Both businesses run on data at large scale. Millions of products, prices that change daily, and decisions driven by what customers do from minute to minute.
What drew me to Overstone was the opposite challenge, helping art be understood and managed as a financial asset. In art, every object is unique, the same work may not come back to market for years, and the data is scattered, inconsistent and hard to compare. Yet banks, insurers and wealth managers need the same rigour they apply to any other asset. Applying data thinking to a market like that was a puzzle I could not resist. Since joining, I have worked across the full data pipeline, from extracting and structuring auction and market data from a wide range of sources, to the datasets behind our valuations, our liquidity and risk scores, our quarterly collection reviews and our ongoing monitoring for clients.
That experience has taught me one principle above all. Every figure we give a client should be traceable back to the artwork, the sale or the source it came from.
Overstone combines human expertise with machine learning. From your perspective, what are the biggest benefits and challenges of this human-in-the-loop approach, and how does it enhance the reliability of your predictions?
I work close to the point where the model and our specialists meet, so I see this every day. The model’s strength is scale and consistency. It can score thousands of works in a collection the same way every time and flag what has changed since the last review. Our art specialists bring what the data cannot capture: a reattribution, a work that is “circle of” rather than by the artist, a condition issue, or a sale result that does not reflect true demand.
The biggest challenge is getting those two perspectives to inform each other. When a specialist disagrees with a model output, the most valuable thing is not just the corrected number but the reason behind it. Recording that reasoning in a structured way takes discipline. It is what turns a one-off fix into something the model learns from.
That is how this approach makes our predictions more reliable. Each review is more than a quality check on today’s output, it also becomes a better training example for tomorrow’s. Over time, the model handles more of the routine cases and our specialists can focus on the works that genuinely need an expert eye.
Generative AI and large language models are increasingly shaping the art and data landscape. How do you see LLMs being further integrated into Overstone’s workflow, and what potential do you think they hold for transforming data analysis in the art market?
I see three practical areas where LLMs can make a real difference.
The first is data extraction. Much of my work involves collecting data from sources that were never designed to be analyzed: auction results, catalogue entries, provenance records and condition reports, each in its own format. Traditional extraction logic breaks whenever a format changes, but LLMs can read the material much as a person would and turn it into clean, structured data automatically. That means faster data collection, quicker due diligence, and more time spent on analysis rather than preparation.
The second is access. This year we started internal sessions introducing colleagues to AI tools, and the most encouraging response came from people outside the data team. When anyone can ask a question of a dataset in plain English, data stops being a bottleneck and becomes part of everyone’s work.
The third is reporting. LLMs can draft first versions of collection-review commentary directly from the underlying numbers, so the team spends its time on judgement rather than formatting.
In every case, the rule is the same: the person using the output is accountable for it. LLMs are excellent assistants, but in a market built on trust, every output still needs to be checked against its source.
...and finally, if any artwork could hang on your wall, which one would you choose and why?
I would love two works side by side, because together they say something about my role.
The first is a Bindu painting by S. H. Raza. A founding member of the Progressive Artists Group, Raza spent decades in France refining a single motif, the dot, into a whole visual language of circles, triangles and bands of colour. Working with data every day, I am drawn to how structured those compositions are, and how much feeling he gets out of such simple geometry.
The second is a landscape by Damerla Rama Rao, ideally his painting of the Godavari flowing through the Eastern Ghats. He was born in Rajahmundry in 1897, in the East Godavari region my family comes from, and is often counted among India’s most celebrated painters after Raja Ravi Varma.
The pairing is also a lesson from data. Raza is one of the most widely traded Indian modernists, with a deep auction record behind him. Damerla’s surviving work sits almost entirely in his memorial gallery in Rajahmundry and rarely, if ever, comes to market. One artist is rich in data and the other has almost none. It is a good reminder that a model can only be as good as the market it can see, and that some of the most meaningful art sits outside the numbers.