Work · Ackren engine
Research engineer, parsing
Ackren is an eight-stage deterministic pipeline that turns input into an answer and a derivation log, with no statistical inference anywhere in it. This role works on the parsing stages and on the measured quality of what comes out.
The work
What you would do
- 01
Work the pipeline
The engine composes as a fixed sequence of typed stages. You work on parsing and the stages around it, keeping every step deterministic and re-derivable by hand.
- 02
Raise the measured numbers
Trigram hit rate is 0.461 on the held-out set, up from 0.449. Pruning at a count of two or more removes 55.1 per cent of trigram mass, so a sixteen-megabyte build is a bigram model with trigram exceptions. Move these numbers, and report them with their basis.
- 03
Account for failure, do not estimate it
Silent lookup errors run at roughly six thousand one hundred in a billion queries at the current index width. They are bought off with bits, and the bits cost most of a sixteen-megabyte budget. Count them.
- 04
Keep the log open
The derivation-log schema is versioned and published at v1.0.0. Every answer produces one, and the schema is fixed in the open rather than behind a form.
Requirements
What we are looking for
You have worked on parsing or classical language modelling and you are happy with the maths. Experience with n-gram models, information retrieval, and space-efficient structures such as perfect hashing and fingerprints maps directly onto the work. Above all you measure carefully, and you are willing to publish a failure mode before anyone else finds it.
- NLP and parsing
- N-gram language models
- Information retrieval
- Perfect hashing / fingerprints
- Determinism
- Measurement discipline
Terms
How we work
Fully remote, worldwide, with a few hours of overlap with UK hours, GMT or BST. Corrections are a first-class kind of result here: a withdrawn claim is struck and named, never quietly deleted. Work you produce for Aleten is assigned to Aleten, said here rather than later, and most of the current specification is already public.
What the engine does not solve
Non-conventional implicature is detectable but not resolvable. Long-form coherence is unsolved, because nothing in an n-gram distribution holds an argument across paragraphs. Novel bridging is detected, never resolved, and coverage stays strictly below one. You are not going to close these. You are going to detect them honestly and publish them.
Apply
A short form, read by a person
Show us a parsing, language-modelling or retrieval problem you worked carefully, ideally one where you measured a limit and reported it straight.
The form stores your application in our own database and a person reads it; there is no third-party tracking system. If you would rather write, send a line to work@aleten.com saying you want the research engineer, parsing role, and a person will reply.