Aditya Rathi

Production system, 2023 to present

Models that read job postings and education records

Models that turn job postings and education records into structured data for labor-market research.

I built the pretraining study, the serving path and both education systems. My team and I built the models that cut postings into sections.

  • 8encoder architectures compared in domain-adaptive pretraining
    About this number

    Encoders pretrained on job postings and compared on bits-per-byte. As of Oct 2026.

  • at least 8posting section types the NER model segments
    About this number

    Section labels in the posting segmentation schema. As of Oct 2026.

  • 440 millioneducation predictions on job postings
    About this number

    Education predictions shipped by the postings system. As of Oct 2026.

  • 600 millioneducation records classified, person-level dataset
    About this number

    Education records in the person-level dataset, across university, K-12 and international taxonomies. As of Oct 2026.

Two systems: education in job postings, and education recordsLeft, job postings: my team's models cut each posting into sections, running on a base model I picked and serve; classifiers I built read the education each posting asks for, and their predictions are checked before shipping. Right, a separate system: education records from the person-level dataset go to classifiers I built, which answer only when confident and retrain when the records change. Numbers shown: at least 8 section types; 600 million records; 440 million predictions.Job postingslong and inconsistentEducation recordsperson-level datasetSectionsmy team's models, on thebase I picked and serve8+ section typesEducationclassifiersanswer when confident,retrain on changeEducationclassifierschecked before shippingRecords classified600M recordsPredictions onpostings440M predictions
Two separate systems, one per column.

Why it mattered

Job postings are long and inconsistent, and what a job asks for is buried inside them. Education records arrive in many different taxonomies, with no labels. Both have to become clean categories before the Burning Glass Institute’s research can count them.

What the work does

  1. I trained 8 encoder architectures further on job postings and compared them. The postings models are built on the one I picked.
  2. I built the system that runs those models over every posting within a fixed hardware budget.
  3. My team and I built the models that cut each posting into at least 8 section types, such as education, experience and skills.
  4. I built the classifiers that read what education each posting asks for, and I checked their predictions automatically before they shipped.
  5. I also built a separate system that classifies people’s education records across university, K-12 and international taxonomies. It answers only when it is confident, and it retrains when the records change.

Result

On job postings, 440 million education predictions have shipped. In the person-level dataset, 600 million education records have been classified.

What the extraction finds in a posting

Every posting is cut into the sections the downstream models read: responsibilities, experience, skills, education, certifications. Pick a posting to see where each part lands. My team and I built the models that do this in production.

Illustrative. A person labeled these lightly trimmed excerpts by hand, so the tags are not production model output. Nothing runs in your browser.

Each tag names the section a part belongs to.

Data Scientist Department of the Air Force

Duties

  • responsibilitiesPerform data acquisition, processing, and cleaning of large data sets
  • responsibilitiesLeverage machine learning and other analytical techniques to interpret large amounts of data to gain insight and identify patterns
  • responsibilitiesDeliver written status updates, formal technical reports, and briefings to leaders

Qualifications

  • experienceOne year of specialized experience equivalent to the GS-12 grade level in the Federal service
  • skillsExperience using cybersecurity principles, secure software development / acquisition, data security, risk management processes
  • educationDegree; Mathematics, statistics, computer science, data science or field directly related to the position.
  • educationCombination of education and experience: Courses equivalent to a major field of study (30 semester hours) as shown in paragraph A above, plus additional education or appropriate experience

From USAJOBS announcement USTC-25-12790917-DH (Department of the Air Force, opened Aug 27, 2025). A US government work in the public domain.

Registered Nurse, Geriatrics and Extended Care Department of Veterans Affairs

Duties

  • responsibilitiesWill perform job duties independently and in accordance with established departmental and hospital procedures.
  • responsibilitiesReports to the CLC Nurse Manager and Assistant Nurse Manager, but must work in collaboration with the Charge Nurses, Staff Nurses, Nursing Assistants, Medical Support Assistants, Social Workers, Minimum Data Set Coordinators, Restorative and Recreational care, and Providers of the assigned team.

Basic requirements

  • skillsEnglish Language Proficiency. In accordance with 38 U.S.C. 7403(f), no person shall serve in direct patient care positions unless they are proficient in basic written and spoken English.
  • educationGraduate of a school of professional nursing approved by the appropriate accrediting agency and accredited by one of the following accrediting bodies: The Accreditation Commission for Education in Nursing (ACEN) or The Commission on Collegiate Nursing Education (CCNE).

Licensure

  • certificationsCurrent, full, active, and unrestricted registration as a graduate professional nurse in a State, Territory or Commonwealth (i.e., Puerto Rico) of the United States, or the District of Columbia.

From USAJOBS announcement CBSX-13093858-27-AST (Department of Veterans Affairs, opened Oct 16, 2026). A US government work in the public domain.

Electrician U.S. Coast Guard

Duties

  • responsibilitiesInstalling, modifying, repairing, loading and testing systems, circuits, equipment and controls
  • skillsInterpreting and applying the National Electrical Code, local codes, building plans, blueprints, wiring diagrams
  • responsibilitiesDetermining location of existing electrical devices and service entrances
  • responsibilitiesPlanning and laying out work, tracing hard-to-locate defects or problems
  • responsibilitiesInstalling, modifying, repairing, and maintaining electrical lines, circuits, systems

From USAJOBS announcement 20250204-77510-007774-IM-DE (U.S. Coast Guard, opened Sep 15, 2025). A US government work in the public domain.

IT Specialist (Customer Support) Army National Guard

Duties

  • responsibilities…install, configure, troubleshoot, and provide maintenance and training in response to customer requirements
  • responsibilitiesProvide Tier II support for end-user hardware failures
  • responsibilitiesRe-image customer workstations and resolve complex problems
  • skillsDetect damaged or lost files, optimize disk management, measure system performance, control system security

Certification

  • certificationsMust obtain and maintain Security+ Certification within six (6) months of appointment
  • certificationsIncumbent must complete appropriate training and obtain required certifications IAW DoDI 8140 or DOD 8570.01M.

From USAJOBS announcement KY-13031363-AR-26-164 (Army National Guard, opened Sep 2, 2026). A US government work in the public domain.

What I got wrong

What I assumed going in
That the model that scored best in pretraining would always do best on the real task.
What I'd measure first next time
Which pretraining choices changed results on the real task, before running all eight.
What it can't do yet
Get enough examples of the rarest posting sections; next time I would collect those first.