Practice two
We build AI systems the way we had to build our own.
Not a practice we started because the market got interested. We needed a production AI platform to run our own language business, we built it, and it has been processing real work under real deadlines ever since. That system is the reference for what we build for you.
The reference implementation
A production system, not a portfolio piece.
Every architectural decision in it was forced by something going wrong at three in the morning. That is the experience you are hiring.
- 55.6M
- Words processed
- 153
- Language pairs
- 1,994h
- Audio transcribed
- 98.0%
- Automated QA pass rate
in production, to date
104 target languages
multi-speaker included
12 checks on every job
Production volume by file type
- audio5,911 jobs · 19,121,946 words
- pdf602 jobs · 17,208,161 words
- rtf474 jobs · 2,891,836 words
- docx459 jobs · 1,883,292 words
- batch197 jobs · 917,147 words
- txlf488 jobs · 473,453 words
- sdlxliff206 jobs · 306,976 words
- mqxliff66 jobs · 101,647 words
- xml67 jobs · 15,624 words
- pptx5 jobs · 6,597 words
Live counters, refreshed every 60 seconds. Formats include the awkward ones — bilingual CAT files, RTF, scanned PDFs — because production data is not clean.
Capabilities
Seven things we build, and the problem each one is for.
Document & workflow automation
The intake-to-output loop your staff does by hand, running unattended.
A person downloads a file, retypes it into a system, routes it for approval, and emails the result. It takes days and it is where the errors come from.
15 file formats in production, including the ones nobody supports: bilingual CAT files, RTF, IDML and scanned PDFs.
What we build
- Ingest from email, SFTP, S3, a portal or an API — whatever you already use
- Classification and routing rules that mirror your actual approval chain
- Structured extraction into your database, ERP or case system
- Human-in-the-loop checkpoints only where a human genuinely adds value
Document intelligence & OCR
Paper archives and scanned PDFs turned into structured, searchable, translatable data.
Decades of records exist only as images. Search does not reach them, reporting cannot count them, and nobody can afford to retype them.
17M+ words processed out of PDF alone, with page geometry preserved from the source rectangle rather than guessed from pixels.
What we build
- Dual-backend OCR with automatic fallback when one engine degrades
- Layout reconstruction that keeps tables, images and physical page size
- Field extraction against your schema, with confidence scores
- Output as searchable PDF, Word, or rows in your system of record
Multilingual AI systems
Language capability embedded inside your product, portal or workflow.
Off-the-shelf machine translation gets you to 'roughly understandable'. For a legal notice, a benefits letter or a safety procedure, roughly is a liability.
142 language pairs and 101 target languages running in production today.
What we build
- Translation pipelines with your glossary, memory and style rules enforced
- Language detection and routing for mixed-language inbound content
- Multilingual search, chat and IVR over your own content
- Automatic quality gates that hold a job back instead of shipping it wrong
Speech & audio AI
Recordings, calls and meetings turned into accurate, attributed, multilingual text.
Audio is the least searchable asset an organization owns, and generic transcription collapses the moment there are two speakers, an accent, or a second language.
1,388 hours of audio processed, including multi-speaker and code-switched material.
What we build
- Transcription with speaker diarization and timecodes
- Direct speech-to-target-language translation, not a lossy two-hop
- Subtitle generation with reading-speed constraints
- Redaction of names and identifiers before anything is stored
Evaluation & quality harnesses
The part every AI vendor skips: proving the output is good, continuously.
Most AI projects ship without a way to answer 'is it working?'. Six months later nobody can tell whether a model change helped or quietly broke something.
A 12-step automated QA chain runs on every production job, with a 97.6% pass rate we publish rather than assert.
What we build
- Golden datasets built from your real, hard cases — not synthetic samples
- Automated scoring on a defined error typology with severity weighting
- Regression gates in CI so a model or prompt change cannot ship blind
- Dashboards that give reviewers and executives the same number
Private & on-premise deployment
For data that cannot leave your building, your cloud, or the country.
Student records, PHI and privileged material often cannot be sent to a public model API — which is where most AI pilots die in legal review.
We run our own infrastructure rather than reselling somebody else's, so residency is a deployment choice and not a support ticket.
What we build
- Open-weight models deployed in your VPC or on your hardware
- US-resident processing with documented data flow
- No client data used for training, contractually
- Air-gapped operation where the environment requires it
Integration, APIs & webhooks
It only counts if it reaches the systems your staff already have open.
A brilliant tool nobody opens is a failed project. The work has to arrive where the work already happens.
A public API and webhook dispatcher run in production, keyed and rate-limited per client.
What we build
- REST APIs with per-key rate limiting and audit trails
- Webhooks for job lifecycle events into your queue or workflow tool
- Connectors for CMS, LMS, DMS, case management and ticketing systems
- SSO and role-based access mapped to your org chart
What usually goes wrong
Four ways AI projects fail, and what we do differently.
The pilot that never ships
A demo impresses everyone, then dies in legal or security review because nobody asked where the data was allowed to go before building it.
How we handle it
We establish the data constraint in week one and pick the deployment model around it — including fully private, if that is what the material requires.
Nobody can tell whether it works
The system is live, it produces output, and there is no agreed way to say if the output is good. Six months later a model update quietly degrades it and nobody notices.
How we handle it
We build the evaluation harness before the feature, on your real hard cases, and wire it into CI as a regression gate.
It works, and nobody uses it
A separate portal staff have to remember to open. Adoption flatlines and the project is judged a failure even though the model was fine.
How we handle it
We integrate into the system your team already has open — the ERP, the case system, the LMS, the inbox — rather than adding a destination.
The 80% that costs 100%
The demo handled the clean documents. Production is scans, mixed languages, five-year-old templates and a table that spans three pages.
How we handle it
Ugly inputs are our default assumption. Our production system runs on exactly that material every day.
How an engagement runs
Fixed fee first. Big commitments only after you have seen it work.
Discovery sprint
2 weeks · fixed fee
We work on your real data, not a demo set. You get a written technical plan, a working prototype of the riskiest part, and a build estimate you can take to procurement. If the answer is that AI is the wrong tool here, we write that down too.
- Technical plan
- Prototype on your data
- Fixed build estimate
Build
6–12 weeks typical
Fixed scope, milestone billing, a demo every two weeks. We build against the evaluation harness from day one, so 'done' is a measurement and not a feeling. Deployed to your infrastructure or ours.
- Working system in your environment
- Evaluation harness + baseline
- Documentation and handover
Run
Ongoing or handover
Either we operate it under an SLA with monitoring and on-call, or your team takes it and we stay available for the first two quarters. Both are priced up front. No surprise managed-service lock-in.
- SLA-backed operation
- Monitoring & alerting
- Or clean handover
Where the work runs
Our infrastructure
US-hosted, we operate it, you get an SLA. Fastest path to production.
Your cloud
Deployed into your VPC or tenancy. Your security team owns the perimeter.
On-premise / air-gapped
Open-weight models on your hardware for material that cannot leave the building.
AI solutions
Bring the process that eats the most hours.
Two weeks, fixed fee, on your real data. You leave with a technical plan, a working prototype of the risky part, and a number you can put in a budget request.