Rishikesh – AI agent development, Python, OpenAI

An AI infrastructure platform focused on workflow automation, search systems, agent harnesses, and multi-agent orchestration. The product supported advanced AI workloads and automation use cases, with the project backed by participation in the AWS GAIA program and a multi-million-euro funding round.

• Designed and built the Workflow Miner Agent, the company's keystone product, which autonomously discovered automatable business processes and captured them as reusable Arazzo blueprints.
• Architected an open-source AI agent meta-framework underpinning many Jentic agents.
• Built RAG-based semantic search over thousands of OpenAPI specifications, combining dense and sparse retrieval with reranking.
• Designed and built the evaluation harness used across all AI systems for deterministic and non-deterministic grading.
• Built the core import pipeline for ingesting, cleaning, validating, and self-healing large OpenAPI specifications at scale.
• Led developer advocacy, mentored engineers and interns, and served on the final hiring panel.

An ML and AI platform supporting personalized recommendations, churn prediction, and demand forecasting for a large fashion rental marketplace. The product helped improve customer personalization and demand planning across a catalogue of more than one million items, with a focus on scalable machine learning systems and reliable model delivery.

• Rebuilt the two-tower personalised style recommender end-to-end, including model architecture, feature engineering, and user-preference representations.
• Validated the recommender via A/B testing with measurable CTR improvement.
• Consolidated model-serving infrastructure from AWS, GCP, and Rackspace onto a single GCP stack and improved performance SLAs.
• Redesigned the ML evaluation framework to surface actionable insights into recommendation quality.
• Built a churn prediction system using causal analysis, Shapley-based feature attribution, tree-extraction rule mining, and user clustering.
• Built an anomaly detection system to flag bad actors using behavioural signals, time-series features, and unsupervised clustering.
• Built a multi-signal demand forecasting model and a pricing model balancing revenue with profitability.
• Served on the hiring interview panel.

An AI and data platform supporting multi-agent orchestration, large-scale entity matching, LLM fine-tuning on GPU clusters, and RAG-based QA systems. The product also included full-stack software engineering capabilities for enterprise data management and integration.

• Built a system translating free-text queries into complex canonical queries.
• Maintained GPU clusters and ran extensive fine-tuning of open-weight models using LoRA/PEFT with automated synthetic data generation.
• Migrated monolithic model serving to Vertex AI, resolving critical crashes during traffic spikes.
• Built PySpark pipelines mapping database schema columns to business entities at large scale for M&A data consolidation.
• Created a RAG chatbot over internal documents using LangChain and Pinecone with source attribution.
• Built a two-phase BERT/RoBERTa QA system for product manuals with a feedback-driven fine-tuning loop.
• Built a PySpark recommender that won 2nd prize in a global company-wide hackathon.
• Built reusable synthetic data generation pipelines and an evaluation pipeline to benchmark non-deterministic GenAI models.
• Served on the hiring interview panel.

An enterprise software engineering environment supporting full-stack development, API design, system architecture, and infrastructure engineering. The work covered a broad range of software engineering challenges across application and platform layers.

• Led a 5 member team in the design, development, and deployment of a European Banking application with an existing 3 million user base.
• Re-engineered the applications as RESTful APIs using Java Spring Boot, Microsoft Azure Services, Redis Cache and deployed them on Docker containers for fault tolerance and scalability.
• Designed the NoSQL database using Microsoft DocumentDB for storage and retrieval of application data with millisecond latency and push-button scalability.
• Designed event-driven APIs using service bus and message queues.
• Monitored and cleaned the code regularly for the team and ensured that different components worked and communicated seamlessly.
• Presented design solutions to clients on a weekly basis and communicated design considerations clearly and efficiently.