Open to AI agent & automation engineering roles

I build AI agents that do real work.

Multi-agent sales pipelines, voice agents and RAG help desks — wired into real CRMs and knowledge bases, with a human gate before anything customer-facing. Seven years in regulated IT taught me the rest: trust is the gate.

Explore the engine Ask my AI Claims, with receipts

kelvino-outbound · 9 agents simulated run

Orchestration

agents in one orchestrated sales pipeline

Throughput

28,000+

agent messages across 283 sessions

Deployments

AI systems shipped: sales, voice, help desk

Foundation

7+ yrs

IT in financial services and PHIPA healthcare

Who I am

Seven years keeping regulated systems up. Now I build the agents.

I spent five years in a PHIPA hospital network and then moved to a regulated forex broker, where an outage is a compliance event before it is an inconvenience. That is where I learned the thing this work actually turns on: trust is the gate, not throughput.

Somewhere in the middle of answering the same tickets for the fifth time, I stopped filing them and started building agents that handle them. IRIS came first — a Tier 1 help desk agent nobody asked me for, shipped to 400–500 real users. Then the nine-agent outbound pipeline for Kelvino, then an open-source voice platform, then a Rust pull request to an app I use daily.

I am not a prompt person who picked up operations, and not an operations person who picked up prompts. I write the code, configure the connectors, and stay on call when the thing breaks at 7am. That combination is the whole pitch.

How I work

Four rules

I build it myself. My own prompts, my own skills, my own MCP connectors — not a brief handed to someone else.

Twice means automate. If I do something two times it becomes a skill or a scheduled agent.

Everything leaves a record. Every agent action is logged, and a person approves anything customer-facing.

I learn by shipping. Small end-to-end slices and the docs at every wall. Cloud Run, LiveKit and Rust all arrived that way.

At a glance

open to offers

Based: Barrie, Ontario
Remote, or the Toronto area
Now: IT Support Specialist, OANDA
Regulated forex broker · since Jan 2025
Also: Co-founder, Kelvino.ai
A side venture, with two co-founders
Studied: Honours Computer Programmer Analyst
Georgian College · 4.0 GPA · Dean's List
Certified: ITIL Foundation
IT Service Management · 2021
Writes: Python TypeScript SQL Rust C# Java

Full résumé (PDF)

Flagship architecture

Hermes Agent Engine

An orchestrator and eight specialists: find, enrich, qualify, score and reach out through Twenty CRM — with a human approving every first touch.

Ollama CloudMCPTwenty CRMScrapeGraphAI

Topology · 9 nodes 283 sessions logged

narangi / Routes work between the eight specialists on a Monday-to-Friday schedule, tracks state and reports up to me. one job, narrow toolset

Human approval gate

Nothing customer-facing sends without a person approving the draft

Always on

What it's built to hit design targets

Drop rate

~40%

pruned before scoring

Review time

< 2 min

human, per approved draft

CRM rubric

stages · 0–100 score

Messages

28k+

across 283 sessions

the gate, simplified

@gate(action="first_touch")
def send(lead_id):
    lead = twenty.get(lead_id)
    if lead.score < 70:
        return skip("below threshold")
    if not lead.approved_by:
        return queue_for_human(lead)
    return dispatch(lead)

Controlled testing — not results yet Source

Qualification layer

Five checks before a lead is ever scored

this is where the ~40% goes

Chatbot detection

Already running an AI assistant? Not a buyer.

Review analysis

Missed calls and slow replies in the reviews are the buying signal.

Owner credentials

Is there a real decision-maker to reach at all?

Multi-location

Chains buy differently from a single room. Route them apart.

Aggregator listings

Directories posing as real businesses get dropped outright.

Survivors get scored 0–100 against an ICP built from hard filters, soft filters and qualification criteria — then routed into the 21-stage pipeline.

Production deployments

Three systems in production, one in the open.

Each solved a real problem, used real data, and left a record you can check.

System 01

AGPL-3.0

Kelvino Voice

An AI receptionist that runs on your own machine. Answers calls, texts and web chats 24/7, books appointments, remembers callers and logs to the CRM. About 265 MB, local by default.

Sized to the machine. Qwen 2.5 at 3B on 8 GB, 7B on 16 GB, 14B on 32 GB.
Voice stack. Whisper STT, Kokoro TTS across 54 voices, Silero VAD, LiveKit and Twilio SIP.
Trust built in. Do-not-call enforcement, rate limits, staff takeover mid-thread, AI-graded tests.

Swap any layer: Ollama · Claude · Gemini · Deepgram · ElevenLabs · Cartesia

github.com/patelvraj810/Kelvino.ai

System 02 at OANDA

IRIS · Tier 1 help desk

Nobody asked for it. Supporting 400–500 users at a regulated broker, I kept seeing the same tickets — so I built an agent that answers from our own history, and hands over a complete story when it can't.

RAG over our own history. Jira tickets and Confluence runbooks, indexed through Vertex AI search.
Serverless on GCP. Cloud Run worker, Firestore per-user state, Gemini picked by complexity.
Escalates properly. The user, the issue, what was tried and the relevant knowledge — so Tier 2 never starts from zero.

Google Chat → Cloud Run → Vertex search → Gemini → resolve or escalate

I had never used Cloud Run or Firestore. I built it in small slices and read the docs at every wall.

System 03 upstream PR

WhimprFlow in Rust

My first time in Rust. Forked a Tauri voice-dictation app and added a capsule UI with multi-monitor support and hover interactions, sound feedback and DMG packaging — then sent it all upstream.

Core. Rust and Tauri against macOS native APIs.
Compared the models. Parakeet against Whisper for local and cloud transcription.
Shipped the UI. Floating capsule HUD, multi-monitor coordinates, sound cues, DMG packaging.

Rust · Tauri · Whisper · Parakeet

Merged upstream

Next · built in public The same qualification loop, on HubSpot

5 stages

Stage 01

Capture

Form fires a private-app webhook

Stage 02

Enrich

Firmographics in a waterfall, sources on the timeline

Stage 03

Score

ICP as custom properties; fit sets lifecycle stage

Stage 04

Gate

Agent drafts a task; a rep approves before it sends

Stage 05

Route

Owner by territory, SLA on the MQL→SQL handoff

Claims, with receipts

Anyone can say “AI-native”. Check mine.

Every line here points at something you can open, run or verify.

How I start on a new team

My first 30 days, anywhere.

Wire, ship, compound — the cadence I use on my own builds, applied to your stack and your bottlenecks.

Weeks 1–2Phase 01

Wire and connect

Learn the stack, the buyers and the rules

Learn the products, buyers and internal systems beyond one lane
Encode the ICP and qualification rules as CRM properties and a score
Connect agents to the real stack via MCP or APIs, logging every action

Measurement before anything goes live

Week 3Phase 02

First workflow live

Highest-leverage motion into controlled production

One agentic workflow in production on the highest-leverage motion
Human approval on anything customer-facing
A weekly readout: what it did, what it cost, what changed

Human review under two minutes

Week 4Phase 03

Compound

Make it repeatable, then hand it off

Trigger-based plays on real events: funding, hiring, regulatory change
A reliable handoff to the humans downstream, with SLAs
The playbook documented as reusable skills

One problem outside my lane, off someone's plate

Career

Before the agents, the systems they run on.

Two regulated environments, 400–500 users, and the identity, endpoint and infrastructure stack underneath them.

Jan 2025 – now

OANDA

IT Support Specialist · Toronto

Regulated forex broker · 400–500 users across multiple offices

Microsoft 365 Azure AD GCP SailPoint Jamf JumpCloud ManageEngine Genesys Cloud Salesforce

Built and shipped IRIS. A Tier 1 help desk agent using RAG over Jira tickets and Confluence runbooks, on Cloud Run with Firestore and Google Chat — automating troubleshooting, escalation and knowledge capture.
Designed the escalation path. Workflows that open Jira tickets carrying the user, the issue, what was already tried and the relevant knowledge.
Turned recurring fixes into documentation. An automated workflow that generates Confluence FAQ pages from repeat support interactions.
Tier 1 through Tier 3 in regulated financial services. Microsoft 365 and Azure AD administration across user provisioning, access management and enterprise identity.
Identity, endpoint and call routing. SailPoint access governance, Jamf and JumpCloud device compliance, ManageEngine, Genesys Cloud, Salesforce, Jira and Confluence administration.
Automated the lifecycle. Onboarding, offboarding and account lifecycle workflows via scripting — plus research into enterprise LLM deployment across self-hosted and cloud.

May 2019 – Dec 2024

Muskoka Algonquin Healthcare

IT Technical Support Specialist · Huntsville

PHIPA-regulated hospital network · multi-site

Windows Server Hyper-V Citrix Azure AD VPN / RDP ITIL

Frontline support under ITIL practice. In an environment where downtime touches patient care, not just productivity.
Ran the server estate. Windows Server infrastructure including Hyper-V hosts, patch management and server migrations.
Kept clinical staff connected. VPN, RDP and Citrix remote access environments; Microsoft 365 across Teams, SharePoint and OneDrive.
Owned identity and access. Azure Active Directory — account creation, group policy, access controls and user administration.
PHIPA compliance in practice. System updates, security patches and backups against healthcare security requirements.
Led infrastructure projects. Technology upgrades and deployments across multiple healthcare locations, coordinating vendors and managing procurement.

2018 – 2021

Georgian College

Honours Computer Programmer Analyst

4.0 GPA, Dean's List throughout. Manulife Award for academic excellence in computer studies, Winter 2021.

Software development, systems analysis, business intelligence, application security and data structures — across C#, C++, Java, JavaScript, PHP, SQL, Swift, Kotlin and Arduino, on embedded, desktop, mobile, web and mainframe.

Aug 2021

ITIL Foundation

IT Service Management

The vocabulary behind how I structure incidents, escalation paths and SLAs — which is why the agents I build escalate properly instead of dead-ending.

side venture

Kelvino.ai · Co-founder

AI automation for med spas and salons

Built with two co-founders. It is where the outbound pipeline and the voice platform come from — and it stays a side venture. I am looking for a full-time engineering seat.

Toolbox

Grouped by what it's for.

Highlighted is what I reach for first. The rest I have run in production or shipped with.

Grounded assistant

Ask my AI anything

Ask about the systems I've built, how I work, or what I'm looking for. It answers from a profile I wrote — and it won't invent results.

Answers only from a profile I wrote streaming · /api/chat

Available for full-time roles

Have a workflow that should run itself?

I'm looking for a full-time role where I own the agent system end to end — growth, support, or wherever the next bottleneck is. Remote, or in the Toronto area.

patelvraj810@gmail.com Résumé (PDF)

LinkedIn ↗ · GitHub ↗ · Barrie & Toronto, ON · +1 647 804 2073