Koushik Sivarama Krishnan

Applied ML Engineer | LLM/RAG Systems, NLP, Computer Vision & Speech

Download Resume View Top Projects

Biography

I'm an Applied ML Engineer specializing in LLMs, RAG, and agentic AI systems. I've shipped end-to-end ML workflows, from experimentation and fine-tuning through deployment and monitoring, across insurance, legal tech, and speech AI. I care about the full stack: model performance, MLOps infrastructure, and building AI that actually holds up in production.

Education

Master of Engineering - Applied Data Science, Sept 2023 - August 2025

University of Victoria

Bachelor of Engineering - Computer Science, Aug 2019 - April 2023

Panimalar Engineering College

Learn More ↓

Experience

Co-op Data Scientist, Insurance Corporation of British Columbia (ICBC)

Built internal AI decision-support systems across claims and policy workflows, improving retrieval quality and prioritization for analyst teams.

September 2024 - Present (full-time)

Agentic AI RAG LangChain Milvus Causal Inference

Developed end-to-end implementation of ICBC's in-house agentic AI assistant using Retrieval Augmented Generation (RAG) with LangChain, Milvus, and MCP servers, working closely with senior data scientists to design architecture and validate performance improvements.
Applied causal inference techniques to evaluate the Comprehensive Medical Assessment (CMA) program, guiding stakeholders with data-driven decisions.
Designed a recommendation engine to prioritize insurance claims, blending business logic and data-driven methods to boost decision support and efficiency.

Research Engine Developer, Justice Data and Design Lab

Delivered a faster legal research assistant and analytics pipeline to surface unmet legal needs for policy and access-to-justice stakeholders.

May 2024 - February 2025 (part-time)

RAG NLP Topic Modeling Power BI Legal Analytics

Collaborated with MITACS to leverage data analysis and machine learning, aiming to enhance access to justice through evidence-based opportunities.
Doubled the performance and efficiency of the research engine chatbot using Retrieval Augmented Generation (RAG) to better identify and address unmet legal needs.
Conducted advanced data analysis on large-scale legal text datasets from Reddit and People's Law School using Topic Modeling and Power BI visualizations.

Machine Learning Engineer, SeiSei.ai

Improved voice model quality and scaled audio data operations to accelerate multilingual speech product delivery.

April 2023 - July 2023 (full-time)

Speech ML Model Fine-tuning Data Pipelines TrueFoundry Team Leadership

Fine-tuned the FreeVC voice conversion model, achieving a 15% performance boost on Hindi audio through hyperparameter tuning and preprocessing.
Directed and supervised a team of interns to deploy ML models with Truefoundry, accelerating project delivery and performance targets.
Enhanced data delivery and automated data pipelines for 5000+ hours of audio, improving system performance.

Computer Vision Intern, Drive Analytics

Built production-oriented sports video analytics workflows and contributed computer vision features linked to measurable business gains.

September 2021 - February 2022 (part-time)

Computer Vision AWS Celery RabbitMQ Django

Engineered a Major League Baseball video analytics system deployed on AWS using Celery, RabbitMQ, and Django.
Led a team in image-to-3D model rendering, increasing profits by 10% through innovative vision techniques.
Conducted comparative analysis of object detection models for glove and baseball detection.

Deep Learning Intern, MURF.ai

Developed scalable speech model tooling for voice cloning and transcription improvements in cloud deployment environments.

December 2020 - June 2021 (part-time)

Deep Learning Speech-to-Text Docker AWS SQS Amazon ECS

Implemented a voice cloning application using state-of-the-art deep learning algorithms.
Improved an existing speech-to-text model for better filler word detection.
Created a comprehensive Docker script to automate pipelines and deployed it using AWS SQS and Amazon ECS for scalable performance.

Projects

View all projects on GitHub

Captain Cargo - AI Delivery Tracking System

Problem: Support teams were spending significant time on repetitive delivery status queries that required no human judgment.

Process: Built a voice-first agentic AI assistant workflow using FastAPI, conversational AI via Vapi, and structured order status retrieval backed by a circuit breaker and caching layer.

Why: Chose voice + API orchestration to eliminate manual touchpoints entirely, with production safeguards so the system degrades gracefully when dependencies go down.

Impact: Converted manual status lookups into a fully automated voice flow achieving 94% eval accuracy and 127ms average latency, with 75 tests covering unit, contract, and integration scenarios.

Learn More

NYC 311 Demand Forecasting & Optimization

Problem: Staffing decisions were reactive, creating persistent mismatch between service demand and available teams.

Process: Built a forecasting and optimization pipeline using historical 311 request patterns, time-series modeling, and Bayesian optimization to generate capacity recommendations.

Why: Chose probabilistic optimization over static thresholds to make staffing plans robust under demand uncertainty, not just accurate on average.

Impact: Produced actionable staffing recommendations that reduced both under-allocation risk and resource waste across service demand cycles.

Learn More

GenZ AI Therapist

Problem: Generic chatbots fail to sustain emotionally safe, context-aware conversations, especially for younger users who need consistency, not just answers.

Process: Designed an agentic AI assistant with conversational memory, empathetic tone shaping, RAG-style context retrieval, and guardrail-aware prompting to keep interactions safe and coherent across turns.

Why: Prioritized tone consistency and response safety over feature breadth; in mental wellness, trust is the product.

Impact: Achieved measurably more coherent multi-turn conversations with near-zero guardrail violations, creating an experience closer to supportive dialogue than a FAQ bot.

Learn More

Brain MRI FLAIR Segmentation

Problem: MRI abnormality segmentation workflows were slow and difficult to operationalize end to end.

Process: Built a full MLOps pipeline covering data handling, experiment tracking, MobileNetV3-based segmentation, and deployment-ready model serving.

Why: Chose a lighter backbone to hit deployable inference speeds on CPU-constrained infrastructure without sacrificing segmentation quality.

Impact: Delivered an end-to-end pipeline capable of running inference without GPU dependency, closing the gap between research prototype and clinical deployment readiness.

Learn More

How I Like to Work

From business objective to production reliability - I own the full arc.

1. Problem Framing

Before writing a single line of code, I map success metrics, constraints, and stakeholder tradeoffs to make sure we're solving the right problem.

Business Goal Mapping Metric Definition Risk Analysis

2. Experimentation

I run structured experiments - baselines, ablations, and offline evals - to validate design decisions with evidence before committing to an architecture.

Rapid Prototyping Offline Eval Ablation Studies

3. Productionization

I ship with observability, fallback behavior, and test coverage built in - because a model that can't be monitored or recovered from failure isn't production-ready.

MLOps Monitoring Reliability

Skills

Production tools and libraries used across recent ML and GenAI projects.

ML / LLM

PyTorch TensorFlow Hugging Face LangChain RAG Agentic AI Prompt Engineering Fine-Tuning Model Evaluation RL Agents Computer Vision Speech ML

Data / Backend

Python SQL FastAPI Flask Django Pydantic REST APIs Async Python PostgreSQL MySQL MongoDB Redis

MLOps / LLMOps

Docker Kubernetes AWS Azure GCP MLflow Model Serving A/B Testing Monitoring CI/CD GitHub Actions

Vector / Retrieval

Milvus Vector Search Hybrid Search Embedding Models Retrieval Evaluation Knowledge Bases MCP Servers

Data / Analytics

Pandas NumPy Scikit-learn Power BI Feature Engineering Experiment Design Causal Inference Time Series

Engineering / Quality

Celery RabbitMQ Git Pytest Integration Testing System Design Linux Agile Delivery

Publications

Advancing Ischemic Stroke Diagnosis: A Novel Two-Stage Approach for Blood Clot Origin Identification

Koushik Sivarama Krishnan, PJ Nikesh, S Gnanasekar, Karthik Sivarama Krishnan

arXiv 2023

Cited by 0

Krishnan, K. S., Nikesh, P. J., Gnanasekar, S., & Krishnan, K. S. (2023). Advancing Ischemic Stroke Diagnosis: A Novel Two-Stage Approach for Blood Clot Origin Identification.

ArXiv Scholar