Computer Engineering @ uOttawa. I build failure-aware systems, software designed around the assumption that networks drop, APIs go down, and hardware misbehaves. The interesting engineering isn't the happy path; it's what your system does at 2 a.m. when the uplink dies.

One rule shows up in everything I ship: nothing critical depends on something that can disappear.

AEGIS — edge-AI worksite safety on QNX 8.0 RTOS · C++20 · Raspberry Pi 5
Real-time person + PPE detection where the entire safety loop runs on-device. The emergency voice alert was generated by ElevenLabs, but it's baked into the device at build time, because a safety alarm that needs an API call to speak is not a safety alarm. GPIO and relay fire at SCHED_FIFO priority 30; audio follows at priority 8. The alarm never waits for the voice.

HyperShift — describe infrastructure in plain English, five agents ship it · TypeScript · Turborepo
Architect → Builder → Sentinel → SRE → FinOps, coordinated by an orchestrator with approval gates, rollback, workspace isolation, and an immutable audit log. Autonomy is easy; accountable autonomy is the hard part.

MetaShift — multi-tenant observability with auto-resolution · Node/Express · React · Supabase
Detects conflicts across distributed services, resolves them, and lets you replay any incident. Tenant-scoped Postgres RLS, JWT + API keys, webhooks, audit log, usage metering: the full operational surface, not just the demo path.

Work-Shift — an approval inbox for AI · TypeScript · Fastify · Cloudflare Workers
AI drafts your Slack messages; nothing sends until a human approves. The same human-in-the-loop conviction as HyperShift, applied to the messages you'd actually regret automating.

AeroGuard — SOC incident triage with LLM confidence scoring, MongoDB Atlas vector search, and a 7-test E2E suite, shipped in 24 hours at HackConcordia. Won MLH Best Use of Snowflake API for a dual-store design that keeps analytics from ever contending with real-time reads.

Chronos — digital twin crisis management across power, airspace, transit, and traffic: 14 live APIs over a NATS broker, three decision frameworks compared side by side, and graceful fallbacks so any upstream outage degrades the system instead of killing it. Built as technical lead of 4 at uOttaHack.

NeuroDesk — agentic middleware that turns natural language into audited action: multi-agent orchestration behind budget limits, approval gates, and a full transaction log. Also: Omni-Context OS, a stateful RAG memory layer across Slack, email, and web, Top 9 at Backboard.io × McHacks.

• Design for the outage first. Offline buffering, graceful degradation, and chaos harnesses aren't afterthoughts in my repos, they're in the initial commit plan.
• The cloud makes the device smarter, never dependent. Enrichment, summarisation, and dashboards live upstream; correctness lives on-device.
• Autonomy needs a leash. Every autonomous system I've built has approval gates, audit trails, and rollback. Trust is a feature you engineer.
• Prove it without the hardware. Mock event generators, load testers, and fault injection so the rest of the stack ships while the Pi is still in the mail.

C++20 / QNX for real-time · TypeScript / Node / Next.js 14 for product · Python / FastAPI for agents and data · PostgreSQL + RLS / Supabase / MongoDB Atlas for data · NATS / BullMQ / WebSockets / SSE / Zod for the glue · plus applied LLM work (multi-agent orchestration, RAG, model routing)

📫 developwith.gt@gmail.com · LinkedIn · YouTube · gautamkhosla.com