A machine studying engineer at Cognizant, Raj Bhowmik, who achieved 98% automation of EDI mapping work, shares his strategy to enabling GenAI programs to work with legacy infrastructure.
Enterprise IT departments face an issue: their information is saved in incompatible repositories. SAP HANA runs alongside BigQuery and Azure SQL. EDI information are available dozens of codecs. Buyer information is unfold throughout programs that don’t speak to one another. In response to a McKinsey report, though 23% of organisations are scaling company AI and 39% are experimenting with them, most fail when confronted with this actuality. Raj Bhowmik, Machine Studying Engineer at Cognizant with a Grasp’s in Machine Studying and Knowledge Science, constructed GenAI programs that truly work in these environments: an clever EDI analyzer reaching 98% automation and decreasing weeks of labor to minutes, an Enterprise AI Knowledge Copilot that queries SAP HANA in pure language, and a sentiment evaluation system enhancing picker effectivity by 80 %. With two papers accepted at IEEE conferences (ICNGN 2025 in Singapore and FMLDS 2025) and expertise judging AI hackathons throughout the U.S., he bridges tutorial analysis with manufacturing deployment. This interview examines the technical structure wanted to make AI programs work with legacy enterprise infrastructure and why hybrid approaches outperform pure LLM options.
Raj, your digital information interchange analyser achieved 98% automation, decreasing weeks of labor to a few minutes by mechanically producing mappings, extensible stylesheet language transformations , and eXtensible markup language for X12, EDIFACT, and IDoc codecs. What prompted you to make use of a hybrid strategy as an alternative of a pure giant language mannequin (LLM) or pure rule-based code?
We selected a hybrid strategy as a result of EDI work sits on the intersection of excessive variability and excessive consequence. Deterministic parts present predictable, testable conduct for structural compliance, whereas an LLM helps with interpretation and edge instances the place inflexible logic doesn’t scale. The guiding concept is straightforward: preserve something that have to be right and auditable in a managed layer, and use generative AI the place it provides flexibility—below strict validation and human-in-the-loop safeguards. It additionally helps with long-term maintainability: as an alternative of regularly increasing brittle guidelines, we will adapt to new associate patterns with out redesigning the entire system. General, it’s a practical stability between accuracy, scalability, and governance for production-grade integrations.
After incomes a level in laptop science from one among India’s high universities, a Grasp’s diploma from IUB, and dealing for 4 years as a software program engineer at a number one Indian IT firm, what prompted you to maneuver into machine studying, and the way does your engineering background affect the creation of synthetic intelligence manufacturing programs right now?
Software program engineering taught me that fashions make up 30 per cent of manufacturing programs. The opposite 70 per cent is tooling, error dealing with, fallback logic, and integration. Once I was constructing a sentiment evaluation system for chats between salespeople and prospects in retail, the machine studying mannequin was easy. Many of the effort went into making certain its reliability in manufacturing: dealing with message queue failures, database load, drift monitoring, and easy efficiency degradation. Knowledge scientists usually create notebooks that by no means depart their laptops as a result of they ignore these points. My background in laptop science implies that I consider database efficiency, API design, deployment pipelines, and failure modes as major constraints, not secondary concerns. Techniques should work on Tuesday morning when the infrastructure is overloaded and somebody’s deployment has damaged upstream providers.
You’re employed as an end-to-end GenAI proprietor from drawback definition by deployment. How does this bridge between enterprise groups, information engineers, and stakeholders truly work when most of them don’t perceive the capabilities of LLM?
I act as a two-way translator. Business groups describe issues, corresponding to gradual customer support response instances or prolonged provider onboarding processes. I research precise workflows to seek out bottlenecks, then decide whether or not AI will assist and what strategy is suitable. Not every thing requires GenAI; typically less complicated options work higher. When AI is sensible, I design the structure: database queries, immediate constructions, vector shops, and human involvement factors. Knowledge groups construct pipelines, and engineers deal with deployment. For stakeholders, I set sensible expectations and measure outcomes utilizing enterprise metrics: time financial savings, error discount, and elevated throughput. When presenting demos, my purpose is to indicate concrete worth in numbers, not spectacular terminology. This stage of translation prevents each overpromising and underutilising AI capabilities.
We’re coming into the age of AI brokers, corporations need assistants that may reply questions and full duties throughout instruments like information platforms, CRMs, ticketing programs, and inner wikis. If I needed to construct an enterprise-grade agent, what structure would somebody use to maintain it correct, safe, and dependable?
I’d begin with an agent orchestration layer that breaks a request into steps (plan → act → confirm) and routes every step to the best system. The LLM is greatest used for intent understanding and decisioning, however execution ought to occur through accepted connectors/instruments (APIs, SQL runners, workflow providers) with strict enter/output schemas. Subsequent, I’d add guardrails: permission checks tied to the person’s identification, coverage guidelines (what the agent can/can’t do), and validation that blocks dangerous actions earlier than they run. For reliability, I’d embody retrieval from trusted sources, testable prompts/templates, and automatic checks (schema validation, unit checks for essential workflows, and fallbacks to human evaluation). Lastly, I’d construct in observability—logging, audits, and suggestions loops—so you may measure accuracy, catch failures, and constantly tighten insurance policies because the agent scales.
Two papers have been accepted on the ICNGN 2025 and FMLDS 2025 conferences, held by the IEEE in Singapore and Los Angeles. What analysis questions linking tutorial machine studying and enterprise implementation do they handle?
Each articles discover sensible purposes of machine studying in a company context, documenting fashions that work in manufacturing however should not apparent from theoretical analysis. Educational venues typically underestimate the challenges of commercial implementation. These conferences, the Worldwide Convention on Clever Computing and Subsequent Era Networks and the Worldwide Convention on the Way forward for Machine Studying and Knowledge Science, appeal to practitioners dealing with related challenges: reliability, information high quality, and measuring real-world impression. Worldwide displays introduce you to completely different regulatory environments and technical requirements, that are essential for programs deployed world wide.
Judging synthetic intelligence hackathons throughout the USA and reviewing the technical content material of a ebook. How do these actions affect your work within the manufacturing system?
Judging hackathons permits me to determine new developments a number of months earlier than they’re applied in enterprises — what instruments groups select, what they encounter, what excites them. That is forward-looking info that enables me to foretell the wants of enterprises. Reviewing the technical content material of the ebook allowed me to determine completely different architectural fashions and hone my essential considering abilities with regard to documentation. Evaluating the approaches of others helps to make clear my very own. Each actions require explaining complicated ideas to completely different audiences: hackathon members, authors, and my colleagues. This translation ability is simply as essential as technical depth in company AI. Stakeholders want to grasp what you might be creating, and engineers want the context of constraints. Coaching others in AI use instances forces codifying what truly works versus demo-impressive approaches.
The following step will probably be synthetic intelligence within the subject of ecology. How can the expertise of integrating a number of databases in an enterprise be utilized to fixing issues associated to environmental information?
Environmental issues are related to disordered, distributed information: sensors, satellites, authorities databases, and native measurements. These are integration duties just like these concerned in integrating a number of databases inside an enterprise. Optimising vitality consumption, predicting environmental impression, and monitoring ecosystems — all of this requires programs that mix disparate sources into sensible conclusions. The stakes are larger, and the suggestions cycles are longer than in enterprise purposes — you may’t do A/B testing on actual ecosystems. However the core abilities are transferable: dependable information channels, stakeholder belief, rigorous measurement. I’m curious about purposes that hyperlink environmental information and coverage selections, just like how technical capabilities and enterprise wants are at the moment linked. The purpose is to make complicated programs comprehensible and usable, whether or not they’re company databases or environmental information streams.
