Jigar PatelPrincipal EngineerToronto ON

I build AI systems end to end.

Jigar Patel, black and white portrait
Jigar Patel — Toronto ON

Now

Arclyx, an AI agent platform for teams: one sentence in, a tested agent out, running on chat, schedules, webhooks, and Slack.

Twelve years shipping software as the engineer clients work with directly: I scope the problem, own the architecture, write the code, and ship. Lately that means agent runtimes and evals on one side and Linux installers, Kubernetes, and disk encryption on the other.

Index

6 work entries · ~1,950 commits · 296 merged PRs across Akash repositories

  1. 2026 → nowLive

    Arclyx

    AI agent platform for teams. Core engineer on evals, observability, the agent loop and billing.
  2. 2025-10 → 2026-08GA

    Akash Homenode

    Consumer gaming PCs become Kubernetes worker nodes on the decentralized GPU cloud, through a guided installer.
  3. 2024 → nowClient work

    Client AI delivery

    RAG and document Q&A deployed inside the client's environment, plus LLM automation for their own engineering process.
  4. 2024 → 2026Live

    Provider Console

    Praetor's job carried into Akash's official Console: wallet auth, onboarding, dashboards, upgrades, cluster control machine.
  5. 2022 → 2024Acquired

    Praetor

    Co-founded. Turned becoming an Akash provider into a guided browser flow, and was acquired by Akash Network in 2024.
  6. 2019 → 2022Archived

    Commerce & POS platforms

    Shopify-embedded point of sale and a food-ordering app, built with a hardware partner on an Ontario Centres of Excellence project.

Case studies

Two projects in depth

Arclyx

An AI agent platform for teams, built at Arc10: describe a job in one sentence and Arclyx builds the agent, picks its tools, tries it on a dry run, then runs it on chat, schedules, webhooks and Slack, answering from the company’s own documents with citations. It launched as AuraMind and was renamed Arclyx in October 2026.

I built the eval harness that gates every agent change, rebuilt in October 2026 with LLM-judge checks, a human review queue, multi-turn scenarios and run-to-run comparison, and the observability behind every run: detail pages, analytics, LLM-written run stories, and a record of every model request with its purpose and its actual cost, which feeds Stripe usage billing. Underneath sit the agent loop, triggers and the new-agent flow, which I also built.

It is a team build: I am one of four engineers, with over 320 merged pull requests authored since March 2026. The code sandbox, tenant isolation, MCP client and original routing layer are teammates’ work.

Read the full write-up

320+
Merged PRs authored since Mar 2026
LLM judge
Eval checks with a human review queue
Multi-turn
Eval scenarios with a cost cap on every run
Run stories
An LLM-written account of every agent run
Baseline · judge · review · compare
The eval harness every agent change goes through
Every call
Model requests recorded with purpose and cost
Sentry
Errors traced alongside run analytics
Auto mode
An LLM picks the model for each request

Akash Homenode

Homenode turns consumer gaming PCs with RTX 3090, 4090 and 5090 cards into Kubernetes worker nodes on Akash, the decentralized GPU cloud, through a guided installer and automated cluster provisioning. Operators need no Kubernetes expertise. I led it at Akash Network from the first commit in October 2025 through alpha in April 2026 to GA in August 2026 — scope, architecture, delivery sequencing, and the alpha, beta and public rollout.

The Linux path is a hardened Ubuntu autoinstall ISO, rebuilt with NVIDIA firmware in the initramfs and a forked Subiquity. The Windows path is an Electron installer that boots through an auxiliary partition, sidestepping the 100 MB EFI system partition. An installation key becomes an ED25519 device identity, every request is DPoP-signed, and mesh and cluster-join credentials are short-lived and shredded after use.

Two production root causes shaped the architecture: a fleet-wide silent downgrade from TPM2 unlock to a keyfile, which led to a redesign of device unlock that kept the encryption guarantees while supporting non-TPM2 hardware; and reboot-only failures traced to a systemd ordering cycle around tailscaled, whose removal took a 90-second stall out of every boot.

Read the full write-up

3090 / 4090 / 5090
RTX cards turned into Kubernetes nodes
PCR 7 + 8
LUKS bound to TPM2, with recovery-key escrow
100 MB
EFI system partition sidestepped on Windows
ED25519
Device identity, every request DPoP-signed
90 s
Boot stall removed with a systemd ordering cycle
2
Production root causes traced and designed out
Secure Boot
Signed images, disks fully encrypted with LUKS
Subiquity
Forked, NVIDIA firmware in the initramfs

ScopeOwnShip

How the work runs, in that order

Scope

I sit with the client, find the real constraint, and write it down before anyone codes.

Own

One engineer accountable for architecture, delivery, and what happens in production.

Ship

Working software in the client's environment, then the next slice.

About

Jigar Patel is a principal engineer in Toronto, twelve years into shipping software as the engineer clients work with directly. He is a core engineer on Arclyx, Arc10’s AI agent platform for teams, and led Akash Homenode from its first commit to GA. Before that he co-founded Praetor, acquired by Akash Network in 2024.