Machine learning consulting, from New York.
We take language-model work past the prototype and into something that runs every day. When the product needs building too, we build that.
Services
- Language models in production
- Retrieval over your own documents. Agents that call your own systems. Extraction whose output downstream code can actually trust. The demo is rarely the hard part; the version that survives a year of real traffic is.
- Evaluation
- Eval sets built from your traffic and wired into CI, so a regression shows up before a customer finds it. Where judgement is genuinely subjective, human review, kept cheap enough that it keeps running.
- Data pipelines
- Ingestion, entity matching, deduplication, warehouse modelling. Also the monitoring that says a source went quiet, which is the failure that otherwise goes unnoticed for weeks.
- Product engineering
- Native iOS, web front ends, the backends behind them. Built by the same people who did the model work, so nothing is lost in the handoff between the two.
Apps
We publish our own software as well. It is where we try things before they reach a client project.
Short-range rain forecasting. A Live Activity on the lock screen tracks the next band of precipitation as it moves toward you, so the question is answered without opening anything.
iOS. Swift, SwiftUI, WidgetKit, ActivityKit.
Receipt and document capture. Photographed paper becomes structured line items you can export, with the extraction checked against totals before anything is saved.
Web. Vision models, Python backend, React front end.
Visit PennyOCRAbout
LLMA is the trading name of TATSMANY LLC, registered in New York and working remotely with clients elsewhere.
We are small on purpose. Whoever scopes your project is whoever writes the code, so there is no gap between what was promised in the first call and what gets built after it.
Our background is production data systems: pipelines handling millions of records a day, entity matching at scale, and the evaluation work that tells you whether a model change helped or hurt. If a language model is the wrong tool for your problem, you will hear that in the first conversation.