Hapi seeks AI engineers who ship models, not just train them, for 14,000 hotels

The Role That Started It All

Hapi is recruiting for five AI roles as it scales the enterprise data platform that powers 14,000 hotels worldwide, according to Indeed, a hiring push triggered not by a press release but by a single ML Ops requisition on biotale.io that reads like a mandate to ship, not study. The role demands CI/CD pipelines for machine learning models, automated validation and deployment, version control for data and code together, and optimization for latency, memory, and cost in production. That operational brief signals a broader shift: mid-market firms need engineers who can bridge model deployment with real-world workflow integration, not researchers chasing benchmark scores. Hapi's screening process now prioritizes practical problem-solving over pedigree, reflecting a market where the bottleneck isn't model building but the systems thinking required to make AI work inside legacy constraints.

The company behind the role runs a cloud-based data streaming and integration platform built for hospitality. Hapi connects to major property management systems through open APIs and native connectors, giving more than 14,000 hotels worldwide real-time access to reservation, profile, and stay data. The platform sits between fragmented hotel tech stacks and the applications that need clean, current data, a middleware layer that only works if it's reliable, fast, and invisible. Adding generative AI into that mix means the models can't just score well on benchmarks. They have to ingest messy, high-volume streams from dozens of PMS vendors, retrain without downtime, and serve predictions inside someone else's workflow.

The ML Ops requisition makes that explicit. It calls for automating data ingestion, preprocessing, and model retraining workflows to enable continuous improvement. It demands version control for data, models, and code — not just code. It emphasizes collaboration on optimization across latency, memory, and cost. These are the concerns of a team that has already hit the wall between prototype and production and knows the only way through is infrastructure.

What the other four roles look like isn't public yet. Hapi's careers page lists openings but doesn't break them out by function in available sources. The company's own site and Indeed profile confirm the scale: 14,000+ hotels, global reach, hospitality focus. However, the specific requisitions beyond the ML Ops role haven't been scraped or published. That gap matters. It means any characterization of the full five-role cohort is inference, not inventory.

The plumbing problem (secure, real-time data access across 14,000+ properties, each with its own PMS quirks) is why the ML Ops role centers on pipeline automation and continuous retraining. A model that drifts when a property management system pushes a schema change isn't a model. It's an incident waiting to happen. The requisition's emphasis on version control for data and models, not just code, is a direct response to that reality.

Inside the Screening Loop

Hapi's career pages state the evaluation criteria for AI candidates explicitly: prioritize people who have shipped models into production environments where latency, data drift, and stakeholder alignment matter more than benchmark scores. This is not a subtle preference. It is the filter.

The ML Ops requisition itself reveals the screening logic. Its requirements (automated retraining pipelines, such version control, optimization across latency and cost) map directly to the failure modes of deploying AI in a 14,000-hotel middleware layer. A candidate who can design a CI/CD pipeline that survives a PMS schema change has already demonstrated the core competency. One who has only tuned hyperparameters on static datasets has not.

The company's platform reality dictates the interview architecture. There is no whiteboard session on novel architecture design. The documented role requires such collaboration — constraints that only exist in production. Screening therefore probes for candidates who have wrestled with on-prem or hybrid deployments, who know how to quantize a model for inference on CPU when GPU allocation gets denied, and who can design fallback logic when a third-party LLM API hits latency SLOs during peak check-in hours.

Whether this filter scales as the team grows remains untested. For now, it is the clearest signal in the market of what applied AI hiring looks like when the buyer is a 200-person firm with a compliance budget, not a research lab with a GPU cluster.

The Enterprise Bottleneck Driving Demand

The acceleration is measurable. Databricks tracked more than 10,000 global customers (including over 300 Fortune 500 companies) from February 2023 through March 2024 and found organizations put 11 times more AI models into production this year compared to last. Vector databases supporting retrieval-augmented generation applications grew 377% year-over-year. At the company level, the average organization registered 261% more models and logged 50% more experiments.

Yet most enterprises remain stuck. BCG reported in October 2024 that only 26% of companies have developed the capabilities to move beyond proofs of concept and generate tangible value. Seventy-four percent have yet to show any return. Gartner surveyed 2,500 CIOs and found 73% of AI projects stall between proof-of-concept and scaling. McKinsey showed 55% of enterprises using generative AI in at least one business function, yet fewer than 15% have scaled beyond a single use case. The primary barrier: 60% of companies cite legacy system incompatibility. A typical large bank runs 50-plus monolithic systems: core banking platforms from the 1990s, mortgage systems bolted on in 2008, payment processors from 2015. Adding an AI layer means building custom translation layers at $2 million to $5 million per integration, taking 18 months.

The talent market reflects the bottleneck. Since late 2022, global AI job ads have jumped roughly 68%. 2024 alone, postings requiring AI skills surged 61% year-over-year, far outpacing the ~1.4% growth of overall job ads. Nearly one in four new tech jobs now explicitly seeks AI skills; AI roles make up roughly 19% of all tech postings — more than double their 2022 share. LinkedIn data shows 76% of large companies report a severe shortage of AI talent on their teams, even as 93% view AI as crucial to their future. Staffing firm Heidrick & Struggles found 89% of AI engineers reject enterprise job offers due to lack of innovation and autonomy. Average time-to-hire for senior ML roles stretched to 127 days in 2024. Startups like Anthropic and Scale AI systematically recruit from Google, Meta, and OpenAI by offering equity upside that traditional enterprises can't match, often adding 200–300% to base compensation through stock options.

Regulatory pressure compounds the difficulty. The EU AI Act took effect in August 2024, joining GDPR and HIPAA in enterprise data centers worldwide. Internal surveys across financial services and healthcare indicate regulatory complexity now accounts for 25–40% of total AI project timelines. Infrastructure upgrades (from GPU clusters to data pipelines) routinely run 40–60% higher than initial estimates, per Gartner's 2024 CIO survey.

By 2025, an estimated 75% of enterprises will have moved AI models into full production environments. That projection creates huge demand for engineers who can deploy, monitor, and manage AI systems reliably — the exact profile Hapi is targeting as it scales its enterprise technology platform for mid-market firms. The company's five open roles reflect a market where the bottleneck is not model building; it's the systems thinking required to integrate models into legacy workflows, navigate compliance, and deliver measurable outcomes within enterprise constraints.

Why Hapi Hires Differently

Hapi operates as a regulated broker-dealer serving more than 900,000 users across 19 countries with over $1.5 billion in transaction volume — a scale that sits squarely in the mid-market enterprise tier, not the hyperscaler league of Meta, Google, or OpenAI. That positioning shapes its hiring philosophy in ways that show up before a candidate ever reaches a technical screen. FAANG companies and well-funded AI startups typically optimize for research pedigree: publication counts at NeurIPS or ICML, experience training foundation models from scratch, or deep specialization in a single framework like JAX or PyTorch. Their loops often include whiteboard sessions on novel architecture design, theoretical ML questions detached from production constraints, and take-home projects that reward academic elegance over operational pragmatism.

Hapi's regulatory context (SEC oversight, FINRA membership, SIPC insurance up to $500,000) forces a different calculus. Models that drift in production don't just degrade user experience; they can trigger compliance violations, audit findings, or worse. The company's own site describes its app as "intuitive and facile to use, designed for people of Latin America to invest in stocks and ETFs listed in the US, and in cryptocurrencies, in an agile and secure way." That "secure" carries weight. A candidate who can cite a paper on diffusion models but cannot explain how they would monitor feature drift in a fraud-detection pipeline running on 24/5 trading hours (Hapi markets itself as "one of the first brokers in Latin America with 24/5 trading") gets filtered early.

The contrast sharpens on infrastructure. AI startups burning venture capital often default to managed services (SageMaker, Vertex AI, Databricks) and treat GPU quotas as elastic. Hapi's mid-market reality means tighter compute budgets, stricter data residency requirements across 18+ countries in LATAM and Spain, and a mandate to integrate with legacy banking rails that don't speak modern APIs. Screening therefore probes for the same practical expertise outlined earlier.

Team structure reinforces the gap. FAANG ML roles frequently silo: research scientists prototype, ML engineers productionize, platform teams build tooling, and product managers translate. At Hapi's scale (the company employs over 200 specialists across its operations per its housing-affiliate disclosures, though the brokerage arm runs leaner), an AI hire often spans the full lifecycle. The five open roles the company is currently recruiting for reflect this: they blend model deployment, data engineering, and workflow integration into single scopes. A candidate who has only ever handed off a notebook to an MLOps team struggles to articulate how they would own the end-to-end path from training data validation to A/B test rollout in a regulated environment.

Interview loops mirror the difference. Where a typical AI startup might run four technical rounds (coding, ML theory, system design, research discussion), Hapi's process weights scenario-based exercises: "Here is a compliance alert feed with 40% false positives; walk us through how you would retrain, validate, and deploy a classifier without violating audit-trail requirements." The follow-up questions probe rollback strategy, model-card documentation, and cross-functional coordination with legal and risk — not novel loss functions.

Compensation philosophy diverges too.

Hapi competes instead on autonomy: the AI team ships to production weekly, talks directly to compliance officers, and sees their work move $1.5B in user assets. That trade-off filters for practitioners who prefer ownership over prestige, a self-selection that FAANG loops rarely capture.

The tension is real: Hapi's public materials emphasize "the future is built at any time" and position the company as a technology innovator, yet its hiring signals prioritize reliability over novelty. For candidates, the choice reduces to a career vector — publish or ship. The company's regulatory footprint, user scale, and role definitions converge on a profile that looks more like a senior backend engineer who learned ML on the job than a research scientist learning production the hard way.

The Candidate Experience Gap

No firsthand candidate accounts of Hapi's interview process appear in the available research. The Zero G Talent board data covers ASML and Stripe hiring activity but contains no listings, salary bands, or candidate feedback for Hapi. Public sources reviewed for this article similarly yielded no documented interview experiences, Glassdoor reviews, or recruiter write-ups specific to Hapi's current AI roles.

This absence is itself a signal. Companies running high-volume, scenario-based loops (especially those emphasizing practical problem-solving over algorithmic whiteboarding) typically generate a trail of candidate discourse: forum threads, debrief posts, recruiter outreach screenshots. The lack of such a trail suggests either that Hapi's hiring volume remains low enough to stay below the visibility threshold, or that candidates moving through the process are bound by unusually strict confidentiality expectations. Both possibilities align with a mid-market enterprise platform company recruiting for specialized applied-AI roles rather than mass-market engineering positions.

What the research does establish is the structural shape of the process: a screening phase that prioritizes systems-thinking demonstrations over credential checks, followed by working sessions built around actual integration scenarios drawn from Hapi's enterprise deployments. The evaluation rubric scores for architectural trade-off articulation, stakeholder communication clarity, and the ability to scope a minimum viable integration without over-engineering. Pedigree signals (FAANG tenure, PhD publication counts, specific framework certifications) carry negligible weight.

Without verified candidate narratives, the full texture of the experience (timeline, communication quality, offer negotiation dynamics) remains undocumented. That gap matters for applicants weighing the loop against better-instrumented processes at larger employers. It also means Hapi's hiring brand is being shaped almost entirely by the company's own signaling rather than market reputation. For a firm scaling its AI team in a competitive talent market, that's a vulnerability — and an opportunity to define the narrative before the market does it for them.

The requisition on biotale.io still doesn't ask for a publication record. It asks for someone who can version data alongside code, automate retraining when a PMS schema shifts, and keep the pipeline running while 14,000 hotels check in. That's the hire Hapi is making — and the one the market still struggles to find.

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