A note on honesty: most of my strongest work is private or client-facing, so I describe those by what they do, not by exposing internals or client names. Where something is public, I link it. Forks and spec-stage work are labeled as such.
Service delivery automation
The deepest work I do is turning recurring service operations into systems that are easier to scope, review, and trust.
Service delivery automation at Viyu
Internal tooling for professional-services and managed-services workflows: SOW intake and generation, project discovery runbooks, PBR/QBR report preparation, invoice-review support, operational evidence packets, and review gates around client-impacting actions. The important pattern is not any single platform. It is making messy service work easier to scope, verify, review, and hand off.
Review-ready operations artifacts
Systems that turn scattered tickets, notes, agreements, usage data, and operational evidence into reviewable artifacts: scope records, business-review packets, audit findings, client-safe clarifications, open questions, assumptions, risks, and next-action lists.
Governed AI-assisted workflows
AI-assisted workflow patterns with human review, evidence trails, preview-before-write gates, and explicit boundaries. I care less about autonomous demos than about systems that can show what they used, what they changed, what they skipped, and where a person needs to approve the next step.
HaloPSA & service-desk AI
A multi-year effort instrumenting a PSA platform with AI triage and public MCP servers.
haloai-ticket-triage-pro
AI triage middleware for HaloPSA: seven expert-role personas analyze each ticket from different angles, suggest priority and category, draft PowerShell/Bash fixes, and write enrichment back to Halo. NestJS, multi-LLM, Next.js PWA.
halopsa-workflows-mcp
An MCP server exposing the HaloPSA Workflows API to Claude and other clients, with auth and token caching. A companion halopsa-tickets-mcp covers tickets; both are listed publicly.
AI agent & MCP infrastructure
The substrate everything else runs on — and where I started early.
Governed agent infrastructure
Private AI infrastructure for governed service-delivery automation: reviewable retrieval, decision support, evidence-backed recommendations, preview-before-write workflows, and clear approval points before client-impacting changes are made.
postgres-mcp-tools
A Postgres + pgvector memory system for AI apps — semantic search and persistent cross-session memory over MCP, with pluggable embeddings.
n8n × AI bridge
Custom n8n nodes for MSP AI workflows plus an MCP server wrapping the full n8n API — so an agent can build and trigger automation pipelines in natural language.
Early agent work
Before "agent" was a product category: an AutoGen-to-Azure SDK integration, an AI-SOAR prototype, and a public aidiscordbot — alongside hands-on forks of MetaGPT, AutoGen, and MemGPT.
Autonomous finance
A five-year arc from a single trading script to a multi-agent system.
finBots
An AI trading platform (paper only). Four analysts debate fundamentals, sentiment, news, and technicals in parallel; a bull/bear round follows; a risk manager and fund manager can veto; a dozen gates run before execution. Python agents, a Fastify API, a Tauri desktop app.
schwab
A personal financial operating system — Schwab + Plaid data, a pgvector "financial memory," a Discord bot as the trading interface, and Grafana monitoring. A CFP-in-a-box experiment.
rhAgentic
A shadow-to-live Robinhood agent with a research-and-safety gate in front of execution. The current edge of the trading work.
Microsoft 365 & Azure tools
Standalone tools that predate and coexist with the platform work.
azure-cost-copilot
A CLI that hunts Azure waste with eight analyzers (idle VMs, zombie resources, oversized SKUs…), confidence-scores each finding, and outputs remediation with effort estimates.
ISRP
Ingests raw M365/Defender security recommendations, risk-scores them by business impact, and generates step-by-step PowerShell remediation for hybrid environments.
teams-3cx-app
A Teams app integrating a 3CX phone system — call control inside Teams and automatic sync of M365 users to extensions.
Real-estate operations
A vertical I automated end-to-end for a property group.
Property finance & compliance
Budget-and-forecasting ETL pipelines (Excel → Postgres → Django, with what-if analysis and approval workflows) and a HUD/NSPIRE inspection processor that turns inspection reports into Microsoft Planner tasks.
Operational glue
Inspection-calendar scheduling, multi-entity payroll processing, Expensify reconciliation, and Planner/To-Do reporting — the unglamorous automation that gives people their afternoons back.
Experiments
Smaller things, kept around because they’re interesting.
Svapna
A "dream engineering" tool exploring targeted memory reactivation across sleep cycles. Exactly as strange as it sounds, and I mean that as a compliment to it.
sully-beal-conjecture
A research platform attacking the Beal Conjecture — parallel counterexample search plus formal proof work in Lean 4. A million-dollar problem I don't expect to solve.
smanji-ancestry
A privacy-first DNA pipeline — parses raw consumer-genetics exports and builds a report entirely on-device, nothing uploaded.
storagemon
A cloud-exposure scanner that finds open buckets, databases, and leaked keys, severity-ranks them, and auto-drafts remediation guides.