Antara Raman Sahay

Multi-Agent Optimization (Reduced execution time by 40%) • Optimized multi-agent system that executes software test cases by interacting with user interfaces in virtual environments. • Identified primary latency bottlenecks through code-level analysis, isolating excessive planner-driven LLM calls as a major overhead. • Deployed a caching-based replay mechanism using the OpenAI SDK and PocketBase, enabling deterministic re-runs by directly invoking sub-agents and tools without planner involvement. • Configured a mode-switching execution strategy (Planner Mode vs RL Mode) with automatic fallback to planner reasoning on failure. • Reduced overall execution time by 40% by eliminating redundant LLM calls and enabling sequential sub-agents and tool-driven execution for repeated execution paths. Customer-Facing Multi-Agent QA Chatbot Suite • Architected a customer-facing multi-agent QA chatbot (AG2) that guided users across the full QA lifecycle and produced production-ready test scenarios, cases, data, and environments. • Created adaptive memory and personalization using Redis, persisted finalized artifacts in pocketbase with hash-based deduplication, and orchestrated multi-modal ingestion (26+ MIME types) via Hatchet workflows and RAG tools. • Productionized the system using ARQ-based guardrails, caching, and retries, enabling reliable, end-to-end test suite generation for beginner to senior QA users.

Automated Code Reviewer • Independently led the end-to-end development of an AI-powered code review system that analyzes pull requests and mimics senior SWE feedback, cutting manual review and release delays by 10+ hours per cycle. • Conducted in-depth research on academic papers, open-source tools, and commercial solutions; identified key gaps and selected optimal modeling strategy. • Fine-tuned the base model and improved performance through quantization and parallel processing, reducing latency by 30%. • Designed and released a CI-ready Python package, a FastAPI model server, and integrated deployment via Bitbucket Pipelines and Jenkins. • Seamlessly operationalized within organizational CI/CD workflows. Smart Drilling Dashboard • Built an NLP-powered dashboard that translates natural language queries into SQL, fetches real-time drilling data from databricks, and computes domain-specific metrics. • Constructed a dynamic graph module using sliding window analytics to detect drilling dysfunctions with embedded domain knowledge. • Generated AI-powered summaries for both data and visual insights; incorporated a memory-based multimodal RAG pipeline (ColPaLi). MDS Data Extraction using Azure AI • Structured and optimized an OCR pipeline to extract structured data from complex PDF documents containing tables and embedded images, using a hybrid approach of traditional OCR (PyTesseract, PDF readers, base64 encoding, etc) and layout-aware models (e.g., CoLPaLi). Evaluated and benchmarked various LLMs and VLMs (GPT-4o, Qwen 2/2.5/2VL-7B, CoLPaLi) for accuracy and reliability using a Human-in-the-Loop (HITL) evaluation framework, structuring outputs to conform to the organization’s Master Data Service (MDS) schema for seamless integration into internal systems. Research & Development • Architected a custom Graph RAG pipeline using domain-specific ontologies for enterprise data retrieval in the oil and gas domain. • Fine-tuned a Small Language Model (SLM) using Supervised Fine-Tuning (SFT) on internal domain data to enhance relevance and performance for enterprise-specific tasks. • Instituted a domain-specific knowledge base using ontology-guided NLP and entity linking to structure insights from drilling reports, completion summaries, and equipment manuals for semantic search and contextual retrieval. LLMOps, MLOps, and DevOps Initiatives • Engineered a Jenkins Shared Library PoC tailored for ML and LLM pipelines, accelerating CI/CD workflows and standardizing deployment across multiple AI services.

Deliver live interactive sessions on core and advanced Data Science topics including Python, SQL, Machine Learning, Deep Learning (NLP, Computer Vision, Time Series), and MLOps.

Researched, conceptualized, and deployed cutting-edge language model architectures (LLMs) to address business challenges using advanced NLP techniques. • Processed and visualized large unstructured datasets, utilizing ML algorithms to extract actionable insights and enhance decision-making efficiency by 25%. • Partnered with product, engineering, and data science teams to deliver scalable AI/ML solutions, reducing processing time by 20% and enhancing model accuracy.

1. Secured scholarship as top 0.1% student of the university for the Student Exchange Program at Livingstone College, USA. 2. President at I-Tech (The Technical Club) 3. served as the Joint - Secretary of the same club 4. Elected thrice consecutively as the CR for my class