$230 to place an engineer, not $3,000

Data Nexus ran the hiring for an engineering staffing operation for nineteen months from 2024. Java engineers, in a continuous flow, placed with client companies that are under NDA and are named nowhere on this site.

There was no recruiter on staff. Interviews were run by the engineers themselves, between their own tasks — six people, an hour or two a day each, against an application flow that was always going to be larger than those hours could hold.

The commercial model is the part worth publishing. Data Nexus hired at junior+ and middle−, a grade below the level the client companies asked for, closed the gap at its own expense over two and a half months of training, and was paid only when the client accepted the person. A model like that lives on two numbers: what it costs to find somebody, and how often the training fails to land.

Both moved. By the end an accepted engineer cost $230 all in, training included, against about $3,000 by the conventional route, and 94% were accepted at the first submission to a client.

At about $3,000 an accepted engineer, hiring at junior+ and middle− for roles written a grade higher is not a strategy, it is a way to lose money slowly. The conventional route to a placement is well understood and its costs are unremarkable: an agency percentage or a recruiter on salary, a screening call for everybody who survives a review of their CV, interviewer hours spent on people a form could have removed, and the candidates who clear the supplier’s own bar and then fail the client’s.

That last group is where the money actually goes, and it is invisible until you divide by it. At an acceptance rate of two per cent, one placed engineer is not one candidate. It is roughly fifty of them, and the forty-nine who did not land carry the whole cost of everything spent on them.

So the arithmetic only closes if the person is already sellable at the moment you find them — which is why every staffing supplier competes for the same small population, and why none of them can afford to train.

The second constraint was interviewer time, and it was harder than the money. Six people who had other jobs could between them absorb a fixed and small number of interviews a week. The schedule tab records 136 interviews across twelve weeks; the flow arriving in the same period was a multiple of it. The question was never how to assess everybody well. It was how to keep away from a human anybody who could be filtered without one.

Both constraints meet at a single decision: whether to spend training money on a particular person. Getting that decision right is the whole engagement, and nothing in a conventional hiring process is shaped to inform it.

Interviewer time was the constraint, not candidate supply

The design follows from naming the scarce resource correctly. Applications were plentiful and free; interviewer attention was neither, and it could not be bought at any price, because the people who had it were shipping software for the same company.

Everything was assembled from what was already on hand. An application form, a scoring sheet in a spreadsheet, and a second tab holding the interview schedule for six people. No procurement, no vendor, no implementation project, nobody employed to maintain it, and the formulas did the arithmetic. The tools cost nothing and the system was live in days — which matters less as a saving than as a fact about sequence: the discipline was designed first and the tooling was whatever could carry it.

Not one screening call in nineteen months

The largest saving in the whole operation is a stage that was removed rather than improved. In a conventional process every candidate who survives the CV review receives a screening call of fifteen or twenty minutes, plus the attempts that fail to connect. Here nobody received one, ever.

Stage	Conventional process	How it ran here
Handling the application	A recruiter reads the CV and decides whether to add it to a call list. Days of silence for the candidate	Automatic approval and an invitation to the assessment in the same minute
Screening	A call of 15–20 minutes, plus the attempts that fail to connect	The stage did not exist
Who performs it	A recruiter on staff, or an agency on a percentage of salary	Nobody. The role was not in the operation
Tooling	An applicant tracking system licensed per seat, an implementation, training the team	A form and a spreadsheet. Nothing spent, live within days
Assessment at interview	An open conversation ending in an impression, not comparable between interviewers	Twenty-two scores on a 0–3 scale, a total and a breakdown across five blocks
What is left afterwards	Correspondence and the memory of the people involved	An archive that answered a new question a year after the operation closed

Seven hundred and seventeen people passed through the funnel and none of them cost a call. At fifteen minutes each, including the attempts that fail to connect, that is around 180 hours, or roughly twenty-two full working days of one person. The fifteen minutes is our assumption rather than a measurement, and it is stated as one: substitute a different figure and the shape of the conclusion does not move.

The saving that does not convert into hours is the one that kept the operation alive. An interviewer never spent an hour to discover after ten minutes that the hour was not worth spending. Across six people for whom hiring was not the job, that is the difference between a programme that ran nineteen months and one that would have stopped in the third.

Twenty-two scores on one scale, and the outcome in the same row

The scoring sheet had five blocks — experience, language core, databases, framework, motivation — and inside them specific questions rather than general topics. Not “level of knowledge” but “these two collection types: internal structure and algorithmic complexity”. Every answer scored 0 to 3, twenty-two scores per candidate, sixty-six maximum, the total computed by formula.

Six different interviewers produced scores that survived being laid side by side. That is a consequence of the wording rather than of training: on a question with a right answer two people diverge markedly less often than on a general impression. The practical effect is that the decision stopped depending on who the candidate happened to get, and a candidate could be handed between interviewers without the assessment starting again.

One column mattered more than the rest, and it looked like housekeeping at the time. The outcome — hired, rejected, withdrew — was recorded in the same row as the score. A conventional hiring sheet holds one or the other: an applicant tracking system holds the decision and throws away the reasoning, and an assessment form holds the reasoning and never learns what happened. What that single column made possible is in the Result below, and it was not the reason the column was added.

An instrument that stores the measurement in one place and the outcome in another cannot be calibrated, ever. Not by a better report, not by a model reading it afterwards. The join has to exist at the moment of writing or it does not exist at all.

The profile decides what training can fix

The score never stood alone. Beside it sat the breakdown across the five blocks, so what a reviewer saw was not “43” but exactly where the person fell away: experience present, databases failed, motivation sound.

In an operation that trains at its own expense, that breakdown is not a nicety, it is the investment decision. An engineer weak on a framework can be hired and brought up in a week. An engineer who is about average everywhere with a hole in motivation cannot be brought up at all, because no curriculum touches it. On a single total those two people are indistinguishable, and one of them consumes the whole $230 and everything after it and returns nothing.

This is the mechanism behind the unit economics of the model, and it is the reason the sheet asked twenty-two questions rather than producing a verdict. A verdict tells you whether to hire. A profile tells you what you would be buying and what it would cost to finish.

The application answered itself in the same minute

An application did not wait for somebody to reach it. Approval and the invitation to sit the assessment went out immediately, automatically, with no human in between.

In a conventional process a queue stands between the application and the first signal from the employer. The CV has to be read, judged and added to a list. At volume that is days, and for all of them the candidate assumes nobody noticed and is talking to other employers. Here there was nothing between the application and the invitation: a person applied and in the same minute received an answer and a task, sat it when it suited them, and arrived at the interviewer already tested.

Three parties gain differently, which is why the change held. The candidate gets an answer instead of silence and can do the work in an evening without taking time off. The interviewer meets somebody who has already passed a check. And the operation stops depending on who happened to have time to work through applications today — the failure mode that turns a hiring push into a stop-start.

Ninety-four per cent accepted at first submission

Of the engineers Data Nexus put forward once training was complete, 94% were accepted by the client at the first submission. Our estimate of the same step by the conventional route is about 2%. That ratio, not the cost of the form or the speed of the reply, is the arithmetic of the whole operation.

At two per cent an accepted engineer is roughly fifty candidates, and each of the forty-nine who did not land still carries a review, a call, interviewer hours and in most cases a fee to whoever sourced them. At ninety-four per cent it is barely more than one. Nothing else about the two routes has to differ for the cost of an accepted engineer to move by better than an order of magnitude, and in this operation nothing else needed to.

The two rates are measured against different flows, and saying so is part of the figure. The 94% is of people submitted to a client after training. The one in five that appears in the funnel above is of everybody who ever completed the form. The first number is what the client experienced; the second is what the operation cost.

What the ratio bought is the licence to hire below the bar. The $230 covers everything, including ten weeks of training before a candidate was ever shown to a client. At $3,000 a head carrying that programme is reckless; at $230 it inverts, and the margin on a placement recovers the investment inside the first month. Every incentive in the model then points at whether the capability actually arrived, because a programme that looked excellent and changed nothing cost Data Nexus the entire investment and returned nothing.

The funnel was an instrument and we did not read it

Because every application from every source passed the same assessment on the same scale, the sources were comparable — not by volume, which says nothing and is trivially inflated, but by what share reaches the assessment, what distribution of scores it produces and how many people end up placed. That is the only honest way to answer which channel is better, and a channel producing half the applications and twice the score distribution is the better buy.

The instrument was built and the reading was never taken. There is no source column in the export. Where the channel should sit there is a link to a CV file: of 437 populated cells, 366 point at file storage and around forty at job-board profiles. The channels cannot be compared from this archive, not then and not now.

The conclusion is plain and it cost eighteen months of data. The mechanism gives you comparability; only a recorded source gives you the metric. This is the failure the buyer’s check on testing a growth number is written against — a figure whose base and definition were never fixed cannot be rescued by anybody’s reporting layer afterwards. The fix is not a discipline, it is a domain invariant: the field is mandatory and stamped by the system that stores the record, not by the person filling in the form.

What a model does in this system now

None of the work above used a model. It was a form and a spreadsheet, and that is worth stating plainly before describing what would be different today, because the sequence is the finding.

Four steps were manual and no longer need to be. Intake and acknowledgement, with the source field stamped on arrival. Reading a CV into the same fixed fields the sheet already asks for, so the interviewer opens a populated profile rather than a blank one. Booking against interviewer calendars with the daily limit enforced inside the flow rather than written in a header and checked by eye. And the weekly answer — where the threshold falls, what share of the flow is lost by our decision and what share by the candidate’s — which was computed by hand, once, a year late.

All four depend on the record underneath being structured, and that is the order most organisations attempt in reverse. One row per candidate, a fixed scale, a closed list of statuses, and the outcome beside the score. Put a capable model on top of free-text interview notes and it will summarise them well and tell you nothing you can act on, because the finding in the Result below is not a summary. It is arithmetic over two columns that had to exist first.

Where a model does score or read a candidate, the acceptance conditions are the same ones Data Nexus publishes for any automated agent: a deterministic layer that holds the veto, and an acceptance procedure run before the thing is trusted with a decision. A scored instrument whose validity has been measured is a different object from a model asked for an opinion, and NeuroFrame is where that argument is set out with its own numbers.

One hundred and forty-six engineers started work, from seven hundred and seventeen candidates, assessed by six people for whom hiring was not the job. Four hundred and twenty-nine reached a full interview; two hundred and eighty-eight were filtered by the form before any interviewer time was spent on them.

Those two hundred and eighty-eight are the result rather than a by-product. Four candidates in ten took up not one minute of interviewer attention, and without that the flow would not have fitted into the available hours at all. The relief was not an optimisation. It was the precondition.

What the archive said a year after it closed

The record was opened again for an unrelated question, and it answered one nobody had asked while it was running. It could do so for exactly one reason: the outcome was sitting beside the score.

Forty-three people withdrew of their own accord — accepted another offer, or went quiet. That is 23% of everybody who reached an outcome. Their median final score is 45. For the people who were hired it is 43, and the middle of the two distributions coincides completely, 39 to 49 in both.

On the score, the two groups are the same population. By grade they are not. The operation hired at junior+ and middle−; the people who withdrew were as a rule middle− and middle+, the upper end of what it took on, and that is what a median of 45 against 43 is describing.

Two things stood between an acceptance and paid work, and neither of them is a score. One was ten weeks of training. The other was a move to another city, which the placement required and which a number of candidates declined outright. Both cost more to somebody holding other offers than to somebody holding none — so the stronger the candidate, the less either was worth tolerating, and the loss concentrated exactly where it hurt most.

The operation was considered a success throughout and did not know this for eighteen months. The archive can say it now only because the outcome sits beside the score. An instrument that keeps the two apart reports a healthy funnel and a good conversion rate the entire time, and no cohort analysis can be run backwards over records that never kept what happened.

A threshold needs three outcomes, not two

The second calculation that only became possible in hindsight is where the cut-off should have sat. A threshold of 38 keeps 79% of the people who were eventually hired and removes 88% of those who were rejected. A threshold of 30 keeps 96% of the hires but lets 44% of the weak candidates through. A threshold of 44 removes almost every weak candidate and loses half the good ones.

So a threshold with two outcomes is the wrong instrument. Below the lower bound, an automatic rejection. Above the upper one, straight through. Between them, a human decides — and that band is the only place where interviewer time is worth spending at all. Data Nexus builds this shape into every selection system it now designs, and it is derived from this archive rather than from a methodology.

What we would do differently

The construction was right. Four things in the execution were not, and they are recorded here because they cost more than any of the wins saved.

The form lived apart from the sheet and answers were carried across by hand, which at volume is a separate job for a separate person. The score is written straight into the candidate record now.

Status and level were typed as free text. Across 451 cells there are 38 distinct wordings of status and 19 spellings of level. Twenty-two cells say only “rejected”, and it can no longer be established who rejected whom. That is a closed list of values now.

The candidate source was never recorded, so channels could not be compared by quality even though the mechanism for it was built and running.

Interviewer limits were written in the header of the sheet and checked by eye, and were breached in 9% of cases. A written constraint with nothing to enforce it is a preference.

What this does not prove

Four things, stated here rather than left for a reader to find.

It is one operation, in one discipline, in one labour market. Software engineers applying for contract placements are an unusually measurable population: the competencies are testable in writing, the candidates are used to being tested, and an assessment can be sat at midnight without anybody’s permission. A role assessed on judgement rather than on knowledge does not decompose into twenty-two scored questions, and this record makes no claim that it does.

The 2% comparator is an estimate, not an instrumented baseline. It is what the conventional route was understood to yield at the time; it carries no confidence interval and nobody sampled it. The 94% beside it is our own count from our own records and was not audited by the client companies who did the accepting.

The two causes named for the withdrawals are reasoned, not measured. The training period and the relocation are what the operation knows stood in the way, and the grades of the people who left fit them. But nobody asked the forty-three, the archive holds no exit reason, and a third cause nobody saw would look identical in this data.

Nothing here measures how those engineers performed after placement. Acceptance by a client is the last event in this archive. Whether the people who were trained held up over a year of work is a question the data cannot reach, and it is the question that would matter most to a buyer.

The retrospective findings are hindsight and are worth exactly what hindsight is worth. The threshold analysis and the withdrawal finding were computed after the fact, against outcomes the process itself produced. They say what the instrument would have supported, not what it did support at the time, and the honest summary of the whole record is that the construction was sound and the operator did not read his own dial for eighteen months.