Noisy data breaks AI. Discipline and iteration build trust.
Over the past 15 months, we have been on a journey to bring meaningful AI applications onto our platform. The first step was to build a customer-facing chatbot. It quickly became clear that general-purpose LLMs were not enough and we needed a retrieval-augmented generation (RAG) model built on our platform-specific knowledge. But building that knowledge base was no small feat as the data had to be cleaned, sanitized and validated.
“AI success begins with trustworthy data, purposeful use cases, and continuous iteration.” ~ Dr Pankaj Dixit Chief Technology Officer & EVP, Government e-Marketplace
The problem is that AI doesn’t correct data—it amplifies it. So, if your input is flawed, your output will be flawed too. After much iteration and user feedback, we reached a 90–95% accuracy level and felt confident moving it from beta to production. Why AI Demands Data Discipline and Iterative Precision Noisy data breaks AI. Discipline and iteration build trust. From there, we expanded into voice-enabled ticketing. This use-case turned out to be far easier, thanks to existing APIs for voice translation and simpler backend integration. Users could now raise support tickets and receive updates across multiple languages through conversational AI.
The next big frontier was to move beyond pre-login information to authenticated, user-specific queries—orders, tenders, bids. We are piloting with security and privacy top of mind. In parallel, we are exploring conversational BI, which remains our most complex and challenging use case. The idea of letting users query large databases in natural language and auto-generate dashboards sounds great, but making it work with thousands of columns, live data and multiple layers of context is incredibly hard. O We are not yet ready for beta—but we are getting closer.
We have also been experimenting with code generation and automated documentation, but those models need strict guardrails and validation before deployment. Through all this, one thing has become clear: the success of AI depends entirely on the quality and readiness of data. Another key takeaway is governance. From hallucinations to data privacy, control is non-negotiable. Explainability is another focus—we want new team members to understand how models work, even after people move on. Building AI is not about plugging in a model—it’s about understanding your data, aligning with business needs, and iterating relentlessly