AI Systems · Architecture · Research

Building AI
that works
in the wild.

I design production-grade AI systems — multi-agent pipelines, RAG architectures, and the infrastructure that holds them together when real users show up. Working across Google ADK, Claude ADK, and LangGraph. Writing and building at the frontier of what's possible.

01 Multi-Agent Systems
Google ADKClaude ADKLangGraph
02 RAG & Knowledge Retrieval
KG-RAGHybrid
03 NLP — Pre-GPT to Present
BERTLLMsText-to-SQL
04 Computer Vision & 3D ML
DetectionSegmentationPoint Cloud
05 MLOps & AI Infrastructure
AzureAWSGCP
06 LLM Evaluation & Reliability
EvalsHITLTest Suites
01 / About

The long way into AI

My path into AI started with classical computer vision and NLP — shipping object detection models, 3D point cloud systems, and BERT-based semantic search in production environments, back when none of that came with a convenient API wrapper.

That foundation shapes how I work today. Years spent in the unglamorous parts of ML — data pipelines, model registries, latency budgets, failure modes — means I know what's inside these systems, not just what they can do when everything goes right.

"I care about AI that ships — observable, reproducible, and built to handle real-world messiness."

Today I work at the frontier of multi-agent AI — designing production systems across Healthcare, Finance, and Telecom using Google ADK, Claude ADK, and LangGraph. I've contributed to internal agent frameworks used across large engineering organisations, and I lead teams through the hardest part: turning a compelling demo into something that holds up when real users and real data arrive.

My longer game is to write, research, and build in public — sharing the patterns, pitfalls, and principles that only come from doing this work for real.

Now
AI Systems Architect
Multi-agent systems, GenAI at enterprise scale, internal AI frameworks — Healthcare, Finance, Telecom
Prior
ML & Computer Vision Engineer
CV, 3D point cloud ML, BERT-era NLP, classical ML — Telecom domain at production scale
Earlier
Software Engineer + Research Intern
Enterprise software (Java), and research experience at a national space agency
Foundations
Computer Science — Theory + Practice
Formal CS grounding across two institutions, bridging research rigour and applied engineering
02 / Expertise

Areas of deep practice

🕸️
Multi-Agent Systems
Designing orchestrated agent pipelines across Google ADK, Claude ADK, and LangGraph — decomposing complex tasks, routing between specialised sub-agents, maintaining coherence across long-horizon workflows. HITL checkpoints, telemetry, and failure recovery built in by default.
Google ADKClaude ADKLangGraphMCPHITLTool Use
🗄️
RAG & Knowledge Retrieval
Beyond naive vector search — hybrid retrieval that routes across Knowledge Graphs and dense embeddings, fusing structured and unstructured evidence for grounded, low-hallucination responses in high-stakes domains.
Knowledge GraphsPineconeWeaviateMilvusQdrantHybrid Search
🧪
LLM Evaluation & Reliability
Building evaluation frameworks that actually catch problems before they reach users — automated test suites for multi-agent POCs, hallucination benchmarks, and structured scoring for complex, open-ended outputs.
LLM EvalsMulti-Agent TestingPrompt EngineeringBenchmarking
🔤
NLP — Pre-GPT to Present
A full NLP arc: semantic search with BERT before it was mainstream, enterprise search systems, text classification, and now LLM-powered pipelines including Text-to-SQL, document understanding, and structured extraction at scale. Both the old way and the new way.
BERTTransformersSemantic SearchText-to-SQLLLMsEmbeddings
👁️
Computer Vision & 3D ML
Pre-LLM depth in visual AI — object detection, segmentation, and 3D point cloud models for real industrial environments. The kind of low-level ML knowledge that makes LLM abstractions legible.
PyTorchObject DetectionSegmentationPoint CloudLiDAR
⚙️
MLOps & AI Infrastructure
End-to-end delivery — from experiment tracking and model registry to CI/CD pipelines and cloud deployment. The plumbing that keeps AI systems running (and observable) after the demo is over.
AzureAWSGCPMLflowDockerGitHub ActionsAzure DevOps
🧭
Technical Leadership
Architecture decisions, team mentorship, client-facing delivery, and the judgement to know when to build vs. buy. Leading AI projects from first whiteboard session to production handoff across Healthcare, Finance, and Telecom.
System DesignTeam MentorshipClient DeliveryPOC → Prod
03 / Systems

How I think in architecture

MULTI-AGENT RAG WITH HITL User Query Orchestrator LangGraph Agent Graph ✋ HITL Checkpoint human approval gate Retrieval Agent Reasoning Agent Synthesis Agent Vector Store Pinecone · Weaviate LLM Azure OpenAI · Claude Response
Multi-Agent RAG with HITL
Orchestrated agent graph with human-in-the-loop approval gates, specialist sub-agents for retrieval, reasoning, and synthesis. The production pattern I use for high-stakes domains.
LangGraphHITLAzure OpenAIPinecone
KNOWLEDGE GRAPH RAG PIPELINE Query + Context Semantic Router classifies intent Knowledge Graph Entities + Relations Vector Store Dense embeddings Fusion + Generate LLM synthesis Grounded Response structured semantic
Knowledge Graph RAG
Hybrid retrieval routing across structured Knowledge Graphs and dense vector stores. Fuses symbolic and semantic evidence for grounded responses. Especially effective in clinical and compliance-heavy domains.
Knowledge GraphWeaviateHybrid RetrievalClaude API
GENAI MLOPS PIPELINE Code Commit GitHub Build Docker CI/CD Evaluate LLM Evals Test Suite Register MLflow Registry Deploy Azure / AWS + Monitor continuous monitoring → retrain trigger
GenAI MLOps Pipeline
End-to-end CI/CD for LLM-based systems — evaluation before every deployment, model versioning, and continuous monitoring with automated retrain triggers in production.
MLflowDockerAzure DevOpsGitHub Actions
AGENTIC FRAMEWORK ARCHITECTURE Agent Runtime Core LangGraph-based orchestration State management + routing Reusable Plugins cross-domain tools HITL Interface human-in-the-loop Telemetry observability built-in Chat Interface conversational UI Platform Layer enterprise AI infra Domain Agents vertical-specific
Agentic Framework Design
Internal framework architecture: LangGraph runtime with pluggable HITL interfaces, cross-domain reusable tools, built-in telemetry, and clean separation between the orchestration core and domain-specific agents.
LangGraphHITLTelemetryPlugin Architecture
04 / Writing

Thinking out loud

Aug 2024
Architecting Large-Scale Search Solutions on GCP: A Technical Deep Dive
Search Systems
Aug 2024
Architecting Large-Scale Search on GCP — Part 1: Overview
GCP · Architecture
Aug 2024
Architecting Large-Scale Search on GCP — Part 2: Data Ingestion & Storage
GCP · Data
Mar 2024
Turbocharging Python for Data Science with Numba
Performance · Python
In progress
Production Multi-Agent Systems with LangGraph: Patterns That Actually Work
Coming soon
In progress
Knowledge Graph RAG vs. Vector RAG: Lessons from High-Stakes Domains
Coming soon
All writing on Medium → More technical deep-dives on the way.

Let's think about hard problems together.

I'm open to conversations about ambitious AI systems, technical architecture, and writing collaborations. If you're working on something interesting — reach out.

Send a message →