Anushka Debnath

T

AI Engineer

Thinking Stack

Feb 2025 - Current (1 year 3 months)

As an AI Engineer, I built and deployed a real-time YOLOv8-based logo detection pipeline for TV broadcast analytics, achieving an impressive 98% accuracy. I designed an automated, cloud-based pipeline utilizing Docker and AWS S3 to ingest, process, and store broadcast data at scale, enabling reliable TRP and viewership analytics by validating detections and reducing errors. My experience at Thinking Stack involved developing AI systems closely related to this role, where I constructed a multi-agent system for healthcare workflows, allowing different agents to manage tasks such as document intake, policy validation, and scheduling, which aligns with the concept of Agentic AI. Additionally, I integrated LLM APIs to enhance reasoning and data

T

AI Engineer

Thinking Stack

Feb 2025 - Current (1 year 3 months)

As an AI Engineer, I developed the Watch-Dog (Distributed Infrastructure Monitor), a lightweight peer-to-peer GPU monitoring service utilizing FastAPI and Asyncio for high-frequency, non-blocking health checks, which significantly reduced downtime. I implemented a MongoDB-backed shared-state locking mechanism using atomic upserts to coordinate crash detection across nodes, effectively preventing duplicate alerts. Additionally, I designed a circuit breaker that checks external connectivity before flagging peers, incorporating a confirmation delay to avoid false positives from transient network issues. My experience at Thinking Stack involved building AI systems where machine learning was integral to the workflow. I worked on extracting infor

T

AI Engineer

Thinking Stack

Feb 2025 - Current (1 year 3 months)

As an AI Engineer at MedIntel, I built a multi-agent AI system to automate hospital workflows, including appointment booking, follow-ups, and document intake via a conversational interface, integrating long-horizon planning and tool use. I designed an OCR with a retrieval-augmented generation (RAG) pipeline to digitize unstructured medical records and extract policy-relevant entities, effectively replacing manual data entry in insurance claim processing. Utilizing Chain of Thought (CoT) architectures, I ensured accurate extraction, which led to a reduction in manual claim review and validation efforts by approximately 73% through automated extraction, cross-policy validation, and agent coordination. Additionally, I leveraged my experience w

A

AI Engineer

AIqwip

Sep 2024 - Dec 2024 (4 months)

As an AI Engineer, I developed a multi-agent RAG chat system for educational purposes, leveraging LangGraph and OpenAI to facilitate domain-specific STEM tutoring by answering questions derived from extensive PDF documents. I built evaluation pipelines to enhance agent reasoning and implemented Retrieval-Augmented Generation (RAG) with vector search and agent memory, achieving an impressive answer accuracy of approximately 95% and long-context retention of around 42%. My efforts led to a significant reduction in manual document search and response time by about 45%, transforming unstructured PDFs into an interactive chat-based interface. Additionally, I utilized Django to integrate a voice agent into the application backend, further enhanci

A

AI Engineer

AIqwip

Sep 2024 - Dec 2024 (4 months)

Pure Voice Chatbot LiveKit Integration | GitHubBuilt a real-time voice chatbot with LiveKit, OpenAI real-time API, and Django, supporting multi-accent speech recognition with approx 92% accuracy.Optimized live audio streaming to achieve sub-500ms latency and reduce query resolution time by around 35%.

About

Detail

Timeline

Résumé