← Back// Projects

Things I've built

AI Copilot Studio

BYOK Chat · Idea Filter · Multi-Provider AI

Privacy-first BYOK workflow lets users plug in their own provider keys and keep prompts, responses, and model usage under their direct control instead of being trapped in a closed AI SaaS.
Combines open-ended AI chat with a guided Idea Filter workflow, so users can move from brainstorming to structured business-idea validation in the same workspace.
Supports provider and model choice across OpenAI, Anthropic, OpenRouter, and Groq paths, giving users flexibility on capability, latency, and cost without rewriting their workflow.
Live model discovery reduces setup mistakes by showing exactly which models are available for a given key, preventing failed requests caused by incorrect model names.
Multi-modal support across chat, vision, image editing, and speech-to-text helps users research, validate, and refine ideas from one interface instead of juggling separate tools.
Transparent session UX with fast resets, assistant switching, and usage metadata helps users compare prompts, control spend, and stay focused on decisions rather than tool friction.
Model Context Protocol (MCP) support lets users connect open-source or self-hosted MCP servers — Grafana, Elasticsearch, GitHub, filesystem, and more — giving the AI real tools to call instead of just conversation.
Multi-turn tool-calling loop works across both Anthropic and OpenAI-compatible providers, so the AI can chain tool calls, read results, and keep reasoning until the task is actually resolved.

MCP Use Cases (Connect Your Own Tools)

AWS learning companion — point AWS Labs' open-source MCP servers at your own account to ask the AI about your VPC setup, cost breakdowns, or service configs.
Homelab observability — connect a self-hosted Grafana or Elasticsearch MCP server and ask plain-language questions about your own dashboards and logs.
Developer workflow — link the GitHub MCP server with your own token, or a local filesystem MCP server, to search issues, PRs, or project files conversationally.
Zero-setup demos — public servers like fetch/web-search or the reference memory server need no account at all, so you can try the flow immediately.

Pre- and Post-Change Automation

Built pre/post-change validation for Cisco NX-OS, IOS and IOS XE — reduced validation time from weeks to 10 minutes while preventing "quiet failures" missed by manual checks.
Automated canonical device snapshots: routing tables, ARP/ND, STP, interface states, error counters, and subnet reachability with fan-out pings.
Aggregated CPU, memory, and interface utilization pre- and post-change, enabling regression detection at scale.
AI-driven diff summarization highlights missing VLANs, neighbor changes, and error growth while preserving full evidence-backed reports.
Expanded coverage to switch RMAs and refreshes, normalising OS/version discrepancies for reliable device-to-device comparisons.

PDF Ingestion → pytest → Allure

Replaced a weeks-long contractor-driven process with a repeatable automated workflow executed in minutes.
PDF ingestion pipeline using prompt chaining + RAG to normalise multi-page documents into structured markdown test specs.
Auto-generates pytest scripts with PyShark and testbed YAML files for scalable, repeatable validation.
Allure-pytest reporting delivers visual, versioned certification reports with full traceability for board-level compliance.
Eliminated contractor costs, standardised certification quality, and accelerated onboarding from weeks to minutes.

LangGraph · FastAPI · Kafka

LangGraph supervisor/sub-agent architecture where a supervisor orchestrates domain-specific agents for monitoring, troubleshooting, and automation.
Dynamic tool invocation and intelligent decision-making — reduced manual troubleshooting by 40%.
Full-stack chatbot (FastAPI, Kafka, MongoDB, React, TypeScript) streamlining cross-team collaboration.
Agentic workflows handle SSH multi-command tasks, upgrade validation, and device searches across ExtraHop, NetBrain, and Cisco ACI.

Neo4j · Airflow · Real-time Impact Analysis

Real-time topology visibility of device interconnections — enables teams to predict impact zones during change management and incident response.
Airflow DAGs schedule daily multi-source data collection, ensuring up-to-date device state visibility.
Progressive ingestion (device-by-device) avoids heavy loads and improves pipeline efficiency.
Neo4j graph database for device relationships — operations teams quickly identify dependency chains and failure points.

Kafka · Airflow · AWS S3 · Grafana

ETL pipelines (Kafka + Airflow) extracting and transforming metrics from Grafana, PostgreSQL, Elasticsearch, and Dynatrace into an AWS S3 data lake.
Real-time monitoring dashboards in Grafana — cut data-access latency by 35%.
AWS Glue Data Catalog + Lake Formation for governance; microservices on OpenShift/Kubernetes with 99.9% uptime.
Enabled enterprise-wide analytics: anomaly detection, proactive infrastructure optimisation, and stakeholder reporting.

LangChain · Kafka · Kubernetes

Decomposed monolithic chatbot into 3 scalable microservices with Kafka event-driven communication.
LangChain + Hugging Face RAG pipeline merging LLM knowledge with internal data sources for domain-specific responses.
MongoDB caching eliminated redundant processing; WebSocket notifications via Webex and Slack for live change updates.
Docker/Kubernetes autoscaling deployment — reduced manual troubleshooting and accelerated change validation for network ops.