I build systems that let a model propose and a machine decide.

The parse above is pre-computed — there is no model call on this page. The validation below it is not: that is validateOrder() running here, the same pure function its unit tests cover. It has no access to the confidence score, which is why editing the price to whatever the customer claimed still gets refused.

-- Was the parse good enough to leave alone?
-- Bucketed by the model's own confidence, so we can see whether that
-- number means anything. The tenant predicate is not optional.
select
  width_bucket(parse_confidence, 0.5, 1.0, 5) as confidence_bucket,
  count(*)                                    as orders,
  round(100.0 * avg(case when needed_edit
                    then 0 else 1 end), 1)    as pct_untouched,
  round(100.0 * avg(case when status = 'confirmed'
                    then 1 else 0 end), 1)    as pct_confirmed
from orders
where merchant_id = current_tenant()
  and created_at >= now() - interval '90 days'
group by 1
order by 1;

Confidence correlates with being left alone, but it flattens badly at the top: a quarter of the model’s most confident parses still needed a human edit. That flattening is the entire argument for the validation gate.

The two chunks below the line score in the low 0.7s and are worthless for this question — they are contract boilerplate that reads like every other clause. That is what similarity measures: whether two passages talk alike, not whether one answers the other. Set the cutoff too low and you retrieve fluent noise, which is worse than retrieving nothing, because it is what the model will confidently cite.

// No library. It is eight lines, and the interesting
// decisions are elsewhere.
function cosine(a: Float32Array, b: Float32Array): number {
  let dot = 0, na = 0, nb = 0
  for (let i = 0; i < a.length; i++) {
    dot += a[i] * b[i]
    na  += a[i] * a[i]
    nb  += b[i] * b[i]
  }
  return dot / (Math.sqrt(na) * Math.sqrt(nb))
}

Switching the definition moves week-4 retention by about nine points without a single number in the underlying data changing. That is the point: “retention” is not a measurement until you say what it counts. Note the cohort sizes too — at n=14 one merchant is seven points, so a week-to-week swing here is one person going quiet, not a trend. Later cohorts holding better than earlier ones is the only part of this chart that is evidence the product changed rather than the customers.

• 2024-01 — now

  Product Management Specialist, Consumer · Jawwal, Ramallah

  Fintech products, fixed and mobile features, campaign delivery.

• 2025-01 — now

  Founder · Mo’een, Ramallah

  In pilot with 8 merchants.

• 2021-09 — 2023-06

  MSc Business Informatics · Corvinus University, Budapest

  Stipendium Hungaricum scholarship.

• 2018-09 — 2021-06

  BSc Computer Science · ELTE, Budapest

• 2022-06 — 2022-09

  Data Engineering Intern · ASAL Technologies, Ramallah

  ETL pipelines, AWS, analytics integration.

• Document to structured data

  A pipeline that turns your PDFs, invoices or messages into validated rows, with a review step for anything it is unsure about.

  Includes the validation layer, not just the extraction. The failure mode you care about is a confident wrong answer, and that is what the gate is for.

  Typically3–4 weeks

• Churn and cohort analysis

  Cohorted retention over your actual data, with the queries handed over so you can re-run them without me.

  You get the SQL and the definitions, not a dashboard you cannot change. The definitions are the deliverable.

  Typically2 weeks

• Arabic and MENA-language AI systems

  Extraction or classification that works on mixed Arabic/English text, evaluated on your data rather than on a benchmark.

  Mixed-script, dialectal and code-switched text breaks most off-the-shelf pipelines. I have shipped this and I know where it breaks.

  Typically4–6 weeks

Describe the problem — I’ll tell you if I’m the wrong person for it.