Felix Lists 29 Open Roles — None Are for AI or Engineering Positions

Felix's Hiring Push: 29 Roles, Two Provinces, One Puzzle

A Canadian healthcare company is quietly building a workforce that looks nothing like the AI startup the job boards suggest.

As of this month, Felix lists 29 open positions on its AshbyHQ careers portal across four areas: pharmacy (18 roles), clinical operations (9), BI & data (1), and marketing (1). Employment types split nearly evenly: 12 full-time, 8 part-time, 9 contract. Geography stretches across two provinces and a national remote pool: nine roles in Mississauga, seven in Port Coquitlam, one in Richmond Hill, one restricted to Ontario and Saskatchewan, and 11 fully remote. On-site and remote arrangements each claim roughly half the listings.

Pharmacy dominates. The 18 openings range from pharmacy assistants in Mississauga and Port Coquitlam, to registered pharmacy technicians, staff pharmacists (with a remote Ontario/Saskatchewan option), a relief pharmacist on-call, and a pharmacy manager in Richmond Hill. Clinical operations contributes nine nurse practitioner contracts, several specializing in menopause, mental health, and weight management, all remote across Canada. A single senior data analyst role sits in BI & data, remote nationwide, while a senior growth marketing specialist rounds out marketing.

Zero G Talent's board data, ingested directly from Felix's postings, confirms the velocity: one new role added in the past seven days, a pharmacy manager in Richmond Hill. The board's salary bands across 19 salaried roles run a range with a median; a warehouse supervisor in Mississauga lists; staff pharmacist roles cluster; relief pharmacist lists.

The composition raises an immediate question. The industry narrative sometimes frames Felix as an AI-driven startup redefining technical screening. The live postings describe a pharmacy and telehealth operator scaling clinical and fulfillment capacity: pharmacists, technicians, nurse practitioners, shipping coordinators, a warehouse supervisor. The senior data analyst and growth marketing roles are the only listings that hint at a technology or data team. Either the engineering and AI hiring lives on a separate pipeline not captured here, or the "AI startup" label reflects aspiration more than current headcount reality.

What the Broader Industry Shift Actually Shows

While Felix's own postings don't show an AI engineering hiring wave, a measurable shift is underway across actual AI-native companies: they are discarding algorithmic puzzles in favor of screens that test systems thinking, infrastructure judgment, and the ability to ship models that hold up in production.

Google's research on production ML systems found that model code represents roughly one-tenth of a real-world system. The remaining 90 percent is infrastructure: data ingestion, feature engineering, training pipelines, deployment, serving, and monitoring. Most organizations treat that infrastructure as an afterthought. Then they discover the model was the easy part, and everything around it determines whether AI actually delivers value or just consumes budget.

Recruiting from Scratch's June 2026 tracking of 215 roles at AI startups founded in the last two years shows 62 percent have either eliminated LeetCode-style questions or significantly reduced their weight. This represents a 35 percent increase over 2023. Nearly half now use a take-home project as the primary technical screen. The catalyst is straightforward: generative AI solves most LeetCode problems in seconds. The proxy that algorithmic puzzles once provided (analytical thinking, persistence, technical fundamentals) has collapsed. Google's internal research had already shown brainteasers hold virtually zero correlation with long-term job performance. A 2020 study from North Carolina State University and Microsoft, published in the Journal of Systems and Software, confirmed technical interviews primarily test performance anxiety, not software-building ability.

What's replacing the old model looks more like the actual work. The Four-Pillar Framework proposed by StartupKit structures this as repository reviews, AI fluency assessments, contextual system design, and paid take-home projects. Instead of asking candidates to optimize graph traversal algorithms, evaluators analyze commit history, pull request discussions, CI/CD configurations, and architectural decision records. AI fluency means epistemological skepticism — the ability to distrust, verify, and correct machine-generated code.

The data on outcomes is compelling. Candidates who excel at take-home projects show a 25 percent higher retention rate past 12 months compared to those hired primarily through LeetCode, and they ramp up faster. A bad hire costs a minimum of 30 percent of first-year salary according to SHRM; for a senior engineer at base, total damage routinely reaches total when recruitment, onboarding, lost productivity, project delays, and severance are counted. The trade-off is real: applied problems require more time from both sides. But they yield better signal, fewer mis-hires, and engineers who feel better about the role match and are more likely to accept offers.

This shift also rewrites who gets through the door. Puzzle interviews create an "algorithmic monoculture," systematically favoring candidates with the most disposable time to practice. Working parents, career changers, and engineers from non-traditional backgrounds are filtered out before their actual skills are ever evaluated. Senior engineers with deep architectural wisdom often underperform in whiteboard settings because they're acutely aware of edge cases, production constraints, and the gap between textbook solutions and real systems — awareness that slows them down under artificial time pressure. Meanwhile, candidates who spent hundreds of hours grinding LeetCode without shipping production code thrive in the theatrical environment. The interview selects for preparation time, not engineering capability.

The market is responding. Top engineers increasingly reject companies that subject them to six rounds of irrelevant whiteboard testing. In competitive markets, the companies that respect candidates' time and evaluate relevant skills win the hiring war.

Why Systems Beat Models in Production

The industry shift toward embodied AI and robotics illustrates the engineering depth now required. PUDU Robotics, whose "One Brain, Multiple Embodiments" architecture lets delivery robots, cleaning robots, and industrial robots share a unified AI foundation, demonstrates the infrastructure burden. Their PuduFM foundation model improves continuously from operational data collected across real-world deployments, creating a closed loop between data, algorithms, and products. That loop doesn't exist without infrastructure that ingests heterogeneous sensor streams, versioned model artifacts, and fleet-wide telemetry at scale.

Commercial pressure reinforces the technical imperative. Robotics customers now evaluate solutions on return on investment, operational efficiency, and long-term value creation — not technical demonstrations. The PUDU CC1 cleaning robot achieved global adoption by addressing practical commercial needs, not by chasing benchmark scores. Its deployments at Carrefour, Walmart, and Edeka required mean-time-between-failures metrics that meet rigorous commercial standards in dynamic brick-and-mortar environments. Achieving that reliability demands engineers who think in systems: redundancy, graceful degradation, observability, and automated rollback — not researchers who optimize loss functions in isolation.

The infrastructure payoff is measurable. Companies using optimized ML infrastructure report significant gains: Miro achieved 80 percent compute cost reduction, and Snorkel AI realized over 40 percent GPU cost savings through better resource orchestration. In a capital-constrained environment, those margins determine runway. The screening practices at serious AI shops test for the engineering judgment that produces such outcomes — capacity planning, workload scheduling, cache invalidation strategies, the unglamorous decisions that compound into production viability.

The paradigm is shifting from prompt engineering to architecture that makes prompting irrelevant. As one industry observer put it on LinkedIn: "The best prompt is the one the system writes for itself. Stop chasing the perfect prompt and start building the architecture that makes it irrelevant." That architecture requires context and control to become ambient — governance built into the developer workflow, semantic metadata flowing into every tool, verifiable data provenance reaching both humans and agents. Hiring for platform engineers who build developer tooling, CLI guardrails, and metadata layers signals investment in that ambient layer.

The next stage of AI competition won't be defined by individual model breakthroughs but by comprehensive capabilities: scalable deployment, engineering excellence, product maturity, and global operational experience. The companies that plan for scale from the beginning, adopting infrastructure-as-code, automated CI/CD, and continuous training pipelines, deploy models in hours rather than weeks. Model fluency is table stakes; systems thinking is the differentiator.

How to Prepare for the Real Screen

The Felix data on the board today is a pharmacy retailer hiring pharmacists, technicians, and warehouse supervisors. The industry trend documented above is where frontier AI shops are heading. Job seekers should prepare for the latter's stated bar — because that bar reflects where the market is moving, while verifying which company they are actually interviewing with.

Start with the infrastructure layer. Every deployed model sits on a stack that includes data pipelines, feature stores, serving infrastructure, monitoring, and rollback mechanisms. Candidates who can speak concretely about building or operating one of those components, not just calling an API, separate quickly. That means shipping a side project that ingests messy data, versions features, serves predictions with latency budgets, and alerts on drift. A notebook that trains a model on clean CSV files does not count. The screen rewards evidence that you have wrestled with the parts that break in production: schema changes, backfill failures, cold-start latency, GPU memory fragmentation.

Next, practice reading unfamiliar codebases under time pressure. The most predictive exercise is not writing a new class from scratch; it is opening a 2,000-line repository you have never seen, identifying the bug that causes silent data corruption in the validation path, and proposing a fix that does not introduce regressions. Open-source projects with active issue queues (vLLM, Ray, Triton, Hugging Face Transformers internals) are better training grounds than algorithm puzzles. Pick one, reproduce a reported bug, write a failing test, submit a PR. The review comments you receive are closer to a real screen than any mock interview.

Third, learn to explain trade-offs without jargon inflation. A screen that values systems thinking will ask: "Why did you choose this serialization format? What happens when the schema evolves? How do you measure the cost of a false negative versus a false positive in this pipeline?" Good answers reference concrete numbers you measured — "we saw 40ms p99 overhead with Protocol Buffers versus 120ms with JSON, and the schema registry caught two breaking changes in staging", not textbook definitions. If you have not measured it, say so and describe the experiment you would run.

Fourth, build a portfolio of operational artifacts, not just model cards. A runbook for a production incident you resolved, a capacity-planning spreadsheet that justified a GPU purchase, a postmortem you wrote that led to a process change: these signal the engineering maturity the screen selects for. They also survive the "show me the code" request better than a Kaggle leaderboard screenshot.

Finally, treat the screen as a two-way filter. A process that demands deep systems reasoning but cannot articulate why that reasoning matters for their product is a process that may over-index on performance theater. Ask the interviewer: "What is the hardest production incident your team faced last quarter, and how did the hiring bar you're applying today help or hurt the resolution?" The answer tells you more about the team than any job description.

The Screen That Wasn't There

The research behind this piece reveals a fundamental tension. The "Felix" described in some industry narratives — an AI startup with 29 open engineering and AI roles, refining a technical screen that prioritizes systems-level problem-solving, does not appear in the available data. What the sources document are three distinct entities operating under the name Felix: a K-pop performer (Felix Yongbok Lee of Stray Kids), a Canadian telehealth and pharmacy platform (felixforyou.ca), and an Australian supply-chain software company (felix.net). The first-party board data from Zero G Talent lists 19 salaried roles with a median band, all in pharmacy, warehousing, and logistics across Ontario and British Columbia: roles such as Pharmacy Manager, Warehouse Supervisor, and multiple Staff Pharmacist positions. No AI, machine-learning, or frontier-engineering openings appear in that feed.

The closest the sources come to a "Felix screen" is a single unattributed line in a countermeasures digest: "Felix screen avoids brainteasers trivia leetcode medium hard" and "What Felix does not assess in candidates communication leadership." Those fragments sit alongside biographical notes about the rapper's disc herniation and a 2025 car accident. They cannot be reliably tied to any of the three real-world Felixes, let alone to a unified technical-screen methodology at an AI startup.

What the research does show, concretely, is what the documented Felixes do evaluate. The telehealth platform emphasizes licensed Canadian practitioners, direct insurance billing, and regulatory compliance; its Trustpilot score (4.8) and App Store rating (4.9) reflect patient-facing service quality. The supply-chain platform highlights vendor prequalification, procurement scheduling, sourcing workflows, and contract repositories; its customer testimonials cite 5,000+ vendors managed, 250+ active users, and sustainability-reporting readiness. Neither company's public materials describe a screening process that filters for distributed-systems debugging, model-deployment pipelines, or GPU-cluster orchestration. The board data's salary bands, topped by a pharmacy-manager role, also sit well below the compensation tiers typical for senior AI engineering talent in North American hubs.

If we treat the industry narrative as a hypothesis rather than a fact, the limits of any screen modeled on the described "Felix" approach become clearer by omission. A process that "avoids brainteasers, trivia, leetcode medium hard" and does not assess "communication leadership" would, by definition, miss: collaborative design judgment, cross-functional stakeholder management, regulatory or safety reasoning (critical in both healthcare and supply-chain contexts), and the ability to articulate trade-offs to non-technical leadership. It would also omit domain-specific fluency (pharmacology for the telehealth Felix, procurement law and ESG reporting for the supply-chain Felix), which the actual hiring data suggests are core to those businesses.

The broader industry shift, away from LeetCode-style screens toward practical, role-specific evaluations, is real and well-documented. But anchoring that shift to a "Felix" that exists only in the narrative risks conflating a genuine trend with a fabricated case study. The companies actually named Felix are hiring for regulated, operations-heavy roles where credential verification, compliance tracking, and domain knowledge are the primary filters. Their screens almost certainly measure those things first. Any extrapolation to AI-startup hiring should be labeled as such, not presented as observed practice.

The pharmacy manager in Richmond Hill starts Monday. The AI infrastructure engineer at the actual frontier startup ships a fix to the serving stack that cuts p99 latency by 40ms. Both are real. Only one of them was ever in the data.

Working in frontier tech? Zero G Talent tracks the openings: see every open Felix role, browse frontier tech jobs, the companies hiring, and the people building the field.