Shren Patel — AI Agent Orchestration Engineer

What I’ve shipped

CentrOS

In productionsince Feb 2026

An orchestration platform that turns messy business workflows into agent systems — agents draft guest replies, process refunds and vendor invoices, and an owner approves anything irreversible.

router → executor → adversary → verifier → owner inbox

23-unit hospitality operation · 3,980+ tests on main · 4 vendor primitives → 1 governed layer

In progressA 60–90s walkthrough video: a ticket arrives, risk is classified, the owner reviews, and the tool fires with an audit trail. Until then, try the live demo.

Read the case study

AXED

IP deal2018 – 2020

A WEBTOON Original I created solo and shipped twice weekly, then closed an IP rights deal with Wattpad WEBTOON Studios for a book and a 2-season animated television adaptation.

40M+ reads · 492K+ subscribers · 2M+ monthly views at peak

Read the case study

Also

4

Guest-triage demo3-stage pipeline · runs in your browser · Infinite-Summoner runPipeline pattern
Core concepts PRDcontracts, gates, lanes · 01-core-concepts.md · private repo, on request
Infinite-Summonermulti-stage LLM forge pipelines · private repo, on request
Word Bubbles Dailya daily word puzzle — swipe through scrambled letter bubbles to spell the hidden words

Experience

Ten-plus years — Coursera, Faire, D2L, WEBTOON, now CentrOS. Each row opens.

Founder CentrOS — AI Agent Orchestration Platform Feb 2026 – Present · Remote · CanadaNow
- Built CentrOS production multi-agent orchestration runtime: typed tool contracts, eval gates, LangFuse traces, HITL approval interrupts — collapsed 4 vendor primitives into 1 governed layer.
- Designed HITL interrupt model: confidence-gated handoff, approval checkpoints before irreversible actions, reversible operations, audit trails, escalation policy — 100% of high-risk tool calls human-reviewed.
- Deployed CentrOS in production at a 23-unit hospitality operation (guest comms, refunds, vendor invoices, compliance) — 3,980+ Vitest tests passing on main, CI-backed live ops since Feb 2026.
5 more highlights
- Designed agent runtime on LangGraph state graphs + Temporal-style durability: durable execution, checkpointing, crash recovery, retries, versioned rollouts — 0 silent run loss across 6 months.
- Built eval harness + LangFuse trace pipeline: golden datasets, task-level scorers, regression suites, prompt versioning, failure taxonomy — caught 22+ silent regressions before release.
- Implemented hybrid RAG pipelines on pgvector + OpenSearch: query rewriting, source ranking, chunk hygiene, freshness boundaries, selective context packing — cut prompt token use 40% with grounding intact.
- Built MCP-style tool registry with scoped OAuth permissions, JSON-schema validation, retry policy, side-effect logging — turned ad-hoc API calls into 100% audited typed agent tool invocations.
- Shipped 5+ reusable platform primitives (connectors, secrets, approval gates, eval hooks, deploy templates) — cut new-workflow setup from weeks to 2 days while preserving runtime + security standards.
Stack TypeScript · Hono · PostgreSQL · Drizzle · Vitest · LLM tool-calling
Vertical Automation & Ops Lead Conscious Creations Sep 2022 – Present · Remote · CanadaNow
- Prototyped accommodations workflows on CentrOS: 3-stage guest-triage pipeline (classify → plan → draft) in Python/FastAPI with HITL approval gates before any irreversible action.
- Designed AI-to-human handoff: shared state contract, full conversation history, durable IDs, operator-readable summaries — 0 lost context across 100% of escalation events in live ops.
Stack Python · FastAPI · LLM tool-calling · workflow automation
Operations & Portfolio Manager Independent — Property Management Sep 2022 – Present · Greater Toronto AreaNow
- Ran ops on 20-unit + 3-unit residential portfolio (intake, scheduling, follow-ups, exceptions) — mapped 12+ recurring workflows informing CentrOS agent-automation vs HITL boundaries.
Independent Engineer / Builder Software, AI automation & creative IP Sep 2022 – Jan 2026 · GTA · Remote
- Built LLM workflow automation since ChatGPT launch (Nov 2022): tool-calling/API automations from 2023; LangGraph multi-agent patterns from 2024; MCP integrations from late 2024 — 3+ years compounding.
- Shipped AXED IP rights deal with Wattpad WEBTOON Studios — book + 2-season animated television adaptation (2025) — plus content automation tooling formalized into CentrOS platform Feb 2026.
Stack Python · TypeScript · Node · LLM tool-calling · MCP · multi-agent orchestration
Senior Frontend Developer Faire Jun 2021 – Aug 2022 · Remote · Canada
- Drove apparel pre-order initiatives at Faire: $5M+ GMV across 3+ international markets over 6-month staged workstreams; React/TypeScript/MobX shipped behind A/B guardrails.
- Drove site-wide perf gains via A/B testing + staged rollouts; same guardrail discipline now applied to agent versioning, shadow traffic, and canary releases at CentrOS — 5+ wins shipped.
Stack React · TypeScript · MobX · Jest · Storybook · A/B testing
Creator, WEBTOON Original — AXED WEBTOON Jun 2018 – Dec 2020 · Remote
- Built Python/JS image-processing + consistency-check pipelines for 445K+ subscriber WEBTOON Original at 2x/week ship cadence — 10% fewer re-uploads, hundreds of operator hours saved.
- Designed queue → tooling → review → ship operator workflow at 2M+ monthly views on WEBTOON Original — precursor to today's HITL agent + human production system at CentrOS.
- Brokered AXED IP rights deal with Wattpad WEBTOON Studios — book + 2-season animated television adaptation releasing 2025 — multi-year multi-format commercial agreement structured solo.
Stack Python · OpenCV · image processing
iOS Mobile & Web Developer Coursera Jan 2017 – Apr 2017 · Mountain View, CA
- Shipped iOS course-download feature in 4 months, lifting course completion 20%+ across Coursera's 24M+ learner base; Swift + offline-sync architecture.
- Built React GUI-creation infrastructure at Coursera; cut new-view setup from days to 1 hour and compounded across 5+ teams as the company standard.
- Built Python search tooling lifting course content discovery 60% (pre-LLM NLP era) — same retrieval problem CentrOS now solves with embeddings, hybrid RAG, and pgvector.
Stack React · Python · iOS SDK · Swift · NLP/search
Earlier engineering roles Desire2Learn (D2L) · SpringBoard Data Management · Minted LLC · TransGaming Inc 2014 – 2016
Software Designer / Web Developer, Desire2Learn (D2L)May 2016 – Aug 2016 · Toronto
- Built D2L flashcard module: auto-question-generation + variable-interval testing — boosted student information retention up to 7× vs baseline review across pilot cohort.
Stack JavaScript · Node.js · React · SQL

Mobile Developer, SpringBoard Data ManagementSep 2015 – Dec 2015 · Mississauga
- Shipped re-hauled iOS app with secure authentication + Objective-C overhaul; cut load times 80% on legacy codebase under a 4-month delivery window.
- Built deductions PDF parser automation cutting client manual processing 90%+ — early business-workflow automation pattern carried forward into the LLM tool-calling era.
Stack iOS SDK · Objective-C · PDF processing · auth systems

Software Developer, Minted LLCAug 2014 – Dec 2014 · San Francisco
- Built Python image-metadata extractor + printer-safe color-space converter on AWS pipeline — 50% reduction in reprint requests across Minted's print catalog.
- Built conversion-loss trackers + shipped 3+ hotfixes during Cyber Monday peak (10x+ traffic spike) — early production-incident response discipline.
Stack Python · image processing · color space conversion · AWS

Graphics & Portability Developer, TransGaming IncJan 2014 – Apr 2014 · Toronto, ON
- Built C++ dynamic libraries emulating Windows on macOS for Cider wrapper engine; shipped 3+ QoL features into Guild Wars 2 + Eve Online (3M+ combined players).
Stack C++ · Objective-C · Mac OSX SDK · Windows SDK
How the AI practice compoundedLLMs, tool-calling, multi-agent orchestration and MCP, formalized into CentrOSNov 2022 – today
1. 2013 – 2022Shipping production software — web, mobile, games, data tooling (Faire, Coursera, D2L, Minted, WEBTOON)
2. Nov 2022LLMs deployed across dev & business workflows (ChatGPT → GPT-4 API)
3. Jun 2023Tool-calling pipelines + custom Python/JS automation runtimes
4. 2024Graph-style multi-agent orchestration, RAG, evals, HITL approvals (LangGraph era)
5. Late 2024MCP tool integrations
6. Feb 2026CentrOS — production AI orchestration platform at a 23-unit hospitality operation · daily live ops, 3,980+ tests on main
The compounding practice is in the runtime: 3+ years of LLM tool-calling, multi-agent workflows, RAG, evals, and HITL approval patterns, formalized into one governed production layer. Pairs with domain experts (legal, accounting, ops) on regulated work so high-stakes decisions always get human sign-off at the approval step.

Education

University of WaterlooBachelor of Software Engineering (BSE), Computer Software EngineeringWaterloo, ON · 2012 – 2018

Skills

Agent systems: Multi-agent workflow design & task decomposition

Tool/function calling, structured outputs, schema-validated arguments

Stateful workflow orchestration, checkpointing, crash recovery, replay

Human-in-the-loop approvals, escalation paths, audit trails, governance

RAG, retrieval pipelines, embeddings, context windows, prompt management

Agent evaluation, tracing, observability, latency & token-cost monitoring

MCP servers, webhooks, OAuth-scoped tool access, third-party API integrations

Router · executor · adversary · verifier patterns; confidence-gated handoff

POCs, demos, eval-gated rollout & production hardening for business workflows
LLM practice: LLMs since Nov 2022 · tool-calling since Jun 2023 · multi-agent, RAG and evals since 2024 · MCP since late 2024
Languages: Python 3 · JavaScript · TypeScript · C / C++ · HTML5 · CSS
Data & cloud: PostgreSQL · pgvector · OpenSearch · GCP (BigQuery, Cloud Run, Vertex AI) · AWS (Lambda, Step Functions, DynamoDB, SQS, EventBridge) · Kubernetes · Docker

Full toolboxlanguages, frameworks, databases, cloud, dev tools, integrations

Languages, familiar: Objective-C · C# · VB / VB.Net · Arduino · Processing · MIPS / ARM assembly · uC++ · GML
Frameworks: React + React Native · MobX + Redux · Jest + Storybook · Node.js + Express · FastAPI + Django + SQLAlchemy · TensorFlow + OpenCV · Hugging Face · Socket.IO + WebGL · Bootstrap + SASS / LESS · Selenium + Cocos2d-x · jQuery + Angular · Android SDK · iOS SDK · Windows SDK · Mac OSX SDK · Cocoa Frameworks
Databases: PostgreSQL · MySQL · MongoDB · SQLite · Vector DBs: pgvector · OpenSearch · FAISS
Cloud: GCP: BigQuery · Cloud Functions · Cloud Run · Vertex AI

AWS: Lambda · Step Functions · DynamoDB · S3 · SQS · EventBridge · API Gateway

Kubernetes · Docker · infrastructure-as-code

Serverless architectures · event-driven pipelines

Heroku · CI/CD deployment
Dev tools: Git + Git-SVN · GitHub + Bitbucket · JIRA + Trello · Unix Terminal · Apache Server · MKS Integrity · RegEx · VSCode · PyCharm · XCode · Eclipse · Netbeans · Sublime Text
Integrations & ops: REST APIs · webhooks · OAuth · Slack · email · calendar · CRM-style workflows · Booking & hospitality / accommodations systems · PDF / document parsing pipelines · Git · CI · Unix · AWS · Heroku
Working style: Forward-deployed mindset: gather requirements, ship POCs, iterate with stakeholders · Strong written communication for runbooks, eval reports & customer-facing demos · Background in high-scale web, mobile, and workflow automation since 2013

Roles I fit

8 roles

Staff AI Engineer · Agent Runtime · AI Platform · AI Architect · Workflow and GTM automation · Agent integrations · Forward-Deployed · Serverless production AI. What I bring to all of them:

Durable runtimesCrash-resumable execution on PostgreSQL: checkpoints, retries, rollback and stop-ship, with an explicit failure taxonomy per stage.Seen in CentrOS: the pipeline
Human approval by designRisk tiers and policy lanes as access boundaries, approval before any side effect, and audit trails kept as state.Seen in CentrOS: the owner inbox
Evals and observabilityRegression-gated eval harnesses and traces over cost, latency and failure modes, tested in CI on every merge.Seen in CentrOS: the adversary
Grounded retrievalRAG with ranking, freshness and source attribution, plus token-cost discipline: model routing, prompt design, caching.
Tool and integration contractsTyped, scoped, idempotent tool calls with audit logs; router → executor patterns; the same primitives across LangGraph, Temporal and MCP.Seen in CentrOS: versus the parts
Forward-deployed deliveryCustomer workflow → production agent: process mapping, gated POC → production, and domain experts at the approval step.Seen in CentrOS: the 23-unit deployment

What I bring, role by role

8 roles

Staff AI EngineerOrchestration-layer thinking for the team: durable multi-step execution, runtime state, traceability and retrieval quality.
- Stateful multi-agent workflows with typed state contracts
- Non-deterministic eval harnesses with regression gates
- RAG pipelines with retrieval ranking and freshness controls
- HITL approval surfaces with audit trails
- LangFuse-style observability over trace + cost + failure modes
Agent RuntimeThe layer beneath agent behaviour most teams underinvest in — state, execution, recovery and production ownership.
- Crash-resumable execution with durable state on PostgreSQL
- Gate runners: retries, checkpoints, rollback, stop-ship
- Scoped tool-execution sandboxing with versioned deploys
- Explicit failure taxonomy per stage — no silent corruption
- Streaming LLM responses with cancellation and timeout policies
AI PlatformThe capability layer that lets multiple teams deploy AI safely without bespoke setup: connectors, execution patterns, governance, tooling.
- Tool registries with typed schemas and scoped permissions
- Policy lanes (auto / confirm / careful) as an access-boundary model
- Audit trails as first-class state, not afterthought logging
- Self-service workflow templates with opinionated guardrails
- CI-backed gate tests validating platform changes per merge
AI ArchitectAmbiguous business processes translated into cloud-native AI systems with tool calling, retrieval, state management and approval flows.
- Reference architectures for agent workflows — blueprints, not one-offs
- Framework-parity thinking: same primitives across LangGraph / Temporal / MCP
- Risk-tiered governance language that stakeholders can use
- Multi-vertical domain experience (hospitality, accommodations, ops)
- Production engineering standards: CI / logging / monitoring / rollback
Workflow / GTM automationAgent systems as operational infrastructure, moving work across Slack, CRMs, email, calendars and internal APIs.
- Router → executor patterns with specialized agents and fallbacks
- MCP / tool-registry contracts with scoped permissions
- RAG pipelines tuned for freshness and source attribution
- Self-service workflow templates with guardrailed config
- Narrow action scopes; approvals before any side effect
Agent integrationsThe boundary where agent systems become real products: integrations, context continuity and reliable handoff between automation and humans.
- Evidence timelines preserved across AI → human handoff
- Context-rich session transfer with no lost history
- Audit-logged tool calls: idempotency, retries, rollback per integration
- Explicit session lifecycle phases with state ownership transitions
- Observability over handoff statistics and API failure patterns
Forward-DeployedCustomer workflow → production agent: POCs, integrations and live deployments, paired with domain experts and built under compliance constraints.
- Active production deployment: 23-unit hospitality vertical under direct management
- POC → prod path through gated expansion, not big-bang launches
- HITL demo surfaces tailored for stakeholder reviews
- Domain-expert pairing (legal / accounting / ops) at confirmation gates
- Domain-specific controls layered onto a shared workflow shell
Serverless production AIPython/Node function calling, multi-agent orchestration and vector RAG on serverless AWS (Lambda · Step Functions · DynamoDB · SQS · EventBridge), with LangFuse traces and synthetic testing.
- Contract / gate state machines (Step Functions–style)
- Tool registries with typed schemas and idempotent activities
- Vector retrieval and streaming-response patterns
- CI-backed contract/gate test suites running on main
- Token-cost discipline: model routing, prompt design, caching

Contact

Canada remote · NA-friendly · EST/PST. Happy to walk through CentrOS architecture, framework parity, or a tailored eval and observability slice for your domain.

Emailpatel.shrenil@gmail.com Email
LinkedInlinkedin.com/in/shrenp (opens in a new tab)
GitHubgithub.com/blopit (opens in a new tab)
Phone+1 613-551-9989
References On requestEngineering (a past senior IC or lead at Faire, Coursera, WEBTOON or D2L), operating (a partner from the 23-unit hospitality operation or Conscious Creations) and commercial (the Wattpad WEBTOON Studios counterparty or a Faire cross-functional partner) — available on request, usually after a first call.