Inside HireFlo: How the Engine Works

Most hiring tools are a stack of one-off interviews glued together with calendar invites. We built HireFlo differently — as an engine. This post explains the three layers that make it run, and the one design decision underneath all of them.

The problem we refused to inherit

Every incumbent assessment platform is fighting the same war right now: candidates use AI, so lock everything down. Proctoring cameras, blocked clipboards, browser jails.

We took the opposite side, loudly: AI tools are allowed in every HireFlo stage — and how a candidate uses them is itself part of the score.

That one stance changes everything downstream. If you're not pretending AI doesn't exist, you can stop building surveillance and start measuring what actually predicts on-the-job performance: how someone scopes an ambiguous problem, debugs unfamiliar code, prompts a copilot well, and knows when not to trust the model. That's signal nobody selling lockdown can produce.

Layer 1: The stage library

At the bottom of HireFlo sits a library of twelve reusable assessment primitives — the atoms every interview process is built from. Each stage has a fixed format, a fixed duration, a rubric shape, and a fixed credit price. A few examples:

• Signal Screen — a two-minute async filter at the top of the funnel.
• Async Work Sample — a realistic, messy task graded by AI against a role-specific rubric. Real inputs are dirty; ours are too.
• Scenario Roleplay — a live voice conversation with an AI playing a stakeholder who behaves like real stakeholders do: contradicts themselves, holds back information, changes their mind.
• Live Build with Copilot Judge — an observed build session where AI assistants are switched on and instrumented. We score the sequence of decisions, not just the final diff.
• Debrief Synthesis — a generated hiring-committee document that includes an explicit dissent field: wherever two stages disagreed about a candidate, the debrief says so and links the tape. No black-box verdicts.

Stages are role-agnostic on purpose. The roleplay engine doesn't know it's interviewing a Forward Deployed Engineer today and an Applied AI Engineer tomorrow — the rubric content is injected per role.

Layer 2: Flow recipes

A flow is how a role turns primitives into a process. A recipe picks stages from the library, orders them, sets a pass gate on each, and injects role-specific rubric content.

Take our Forward Deployed Engineer track. The discriminating stage isn't the expensive live build — it's a 25-minute discovery roleplay near the top of the funnel. The AI plays a non-technical ops director whose "weekly reports are always wrong." The candidate's job is to leave with a confirmed, scoped, buildable spec. Proposing a solution in the first three minutes scores down. So does agreeing to everything. The gate is deliberately hard, because ambiguity tolerance is the thing FDEs actually get hired for.

The Applied AI Engineer track opens differently: a flaky extraction agent, three bug reports, and a task that just says "fix it." The rubric is almost entirely about what the candidate does first. Did they build an eval set before touching anything? Did they get a baseline number? Later stages include six failed production traces to diagnose — one of which is a trap, where the model was right and the eval was wrong. That single trace separates people who've run AI in production from people who've read about it.

Because flows are just recipes over shared primitives, modifiers are cheap: a junior variant drops two stages and adds fundamentals, a staff-level variant adds a strategy memo, a campus-volume variant goes fully async with aggressive gating. Two curated tracks become forty configurations without forty products. Adding a new role track costs a config file, not an engineering quarter.

Layer 3: The credit ledger

Pricing follows the architecture. Every stage has a credit price, and credits burn per completed stage, per candidate. The rules are simple enough to publish:

• Credits burn on completion, not scheduling. A no-show costs you nothing.
• A failed gate means downstream stages never bill. The funnel is the discount — candidates who shouldn't go deep, don't, and you don't pay as if they did.
• Calibration and flow editing are always free.
• Every candidate gets one free re-grade, no questions asked.

Run a hundred applicants through the full FDE track and gating means most of them cost you one or two credits, not a hundred. The full funnel lands around fifteen credits per applicant on average — against a typical agency fee on the same hire, the math isn't close.

Why the dissent field matters most

Any AI assessment tool will eventually be wrong about someone. What decides whether teams keep trusting it is what the tool does about that.

That's why every HireFlo debrief surfaces disagreement instead of averaging it away, links directly to the recording behind each score, and why the free re-grade exists. An engine you can audit is an engine you can defend — to your hiring committee, to a rejected candidate, and to a regulator.

What's next

We're starting with two deeply curated tracks — Forward Deployed Engineer and Applied AI Engineer — because they're the roles where traditional interviews fail hardest. Solutions Architect, AI PM, and ML/MLOps tracks are next in line.

If you're hiring for roles where the job is ambiguity, glue code, and judgment under constraints, HireFlo was built for exactly that.

Want to see a flow run end to end? Get in touch.