Selected work
Systems I've shipped
Agentic AI & Retrieval
RAG pipelines, tool-using agents, and retrieval infra shipped in production
Multimodal Agent Platform
A unifying platform bringing multiple analytical capabilities — scoring, root-cause analysis, screen analysis, and search — under one extensible system.
Multi-Agent Query Routing
Chain-of-thought routing across domain-specific subagents for structured query resolution.
Ambiguous Query Decomposition
AI agents that break down vague search queries and dynamically convert variable fields into structured metadata filters.
Long-Context Video Retrieval
Q&A over multi-hour video via chunked, sentence-level retrieval — exposed through a chat interface and an MCP-based agent.
Incremental Vector Search
Semantic search over a continuously refreshing index — retrieval that stays current without a full rebuild.
Document Semantic Matching
A similarity-matching engine reconciling extracted document data against system records using configurable business rules.
Evaluation & Analytics
Measuring and instrumenting model quality in production
LLM-as-Judge Evaluation
An LLM-based scoring approach benchmarked against golden datasets, with a human-in-the-loop feedback cycle.
Automated Root Cause Analysis
A framework-driven RCA system surfacing systemic patterns across call, chat, and email transcripts.
Behavioral Screen Analytics
An analyzer extracting structured operational data — actions, idle time, errors — from raw screen recordings.
MLOps & Infrastructure
Training, deployment, and serving infrastructure for ML systems
Event-Driven SLA Pipeline
A real-time event pipeline built to hit a strict latency SLA, with idempotent handling and multi-source validation.
LLMOps Training & Deployment Platform
An end-to-end automated platform covering data versioning, training runs, model artifacts, evaluation, and deployment.
Custom AutoML Tooling
An in-house hyperparameter-optimization tool for text classification, cutting exploration time from weeks to days.
Distributed Training Infrastructure
Multi-worker distributed training and pipeline infrastructure built on Kubeflow, with Docker-based deployment.
Containerized ML Deployment
A containerized build and publishing pipeline standardizing how trained models moved from experimentation into scheduled jobs.
Lightweight Model Serving
Backend prediction services built around in-memory model loading, minimizing latency for real-time inference requests.
Classical ML & NLP
Traditional modeling and language processing work
Zero-Shot NER at Scale
Applying large language models for zero-shot named-entity recognition across a wide range of document types.
Object Detection Pipelines
End-to-end object detection pipelines covering data processing, model building, evaluation, and deployment.
Classical Text Extraction
Structured text extraction combining statistical models, neural networks, and hand-tuned pattern matching.
Classical Text Classification at Scale
A comparative approach to document classification spanning traditional ML models and early neural methods across a wide range of document types.
Automated Classification & Reporting
Pattern-based text classification paired with automated evaluation reporting on accuracy, precision, recall, and F1.