Tennessee Labs

Applied AI studies
& implementation

Tennessee, United States

Evidence before deployment.

We run applied AI studies for small companies, then build the systems those studies justify. Two domains: medical devices and natural language.

A study ends in a written recommendation. Sometimes the recommendation is not to build.

Fig. 1 Catheter tip position estimated from intracavitary ECG during a simulated advance

20% advanced

Proximal

0.41

60% advanced

Approaching

0.68

85% advanced

Cavoatrial junction

0.94

Motion artifact

No call

—
P-wave amplitude rises as the tip nears the cavoatrial junction, and the estimate reads that feature. Confidence is calibrated against held-out recordings rather than reported raw, so 0.94 has to mean what it says. In the fourth panel motion artifact leaves the P wave unresolvable and the model declines to call it — the behavior that matters most once a device is in someone's hands. Signals are synthetic and illustrative; this is not the output of a validated system.

Scope

Three ways an engagement starts. Most begin at the first.

Feasibility study

Four to six weeks against your data, not a demo set, answering one question: does the approach hold up? You receive a written report — data assessment, baseline results, failure modes, what it costs to run in production, and a recommendation you can act on either way.

Implementation

Design, build, and hand over the working system, together with the evaluation harness that proves it still behaves after we leave. Your engineers own the code and the tests. We document what we did and why.

Validation support

Test protocols, model documentation, and traceable evaluation records that fit the quality system you already maintain. We support your regulatory and quality teams with engineering evidence. We do not replace them, and we do not make submissions on your behalf.

Domains

We stay narrow on purpose. Depth in two fields beats a menu.

Medical devices

Software in and around regulated products: on-device signal and image processing, algorithm performance testing, retrospective evaluation against labeled data, drift monitoring after release, and the documentation that has to survive review. We work alongside manufacturers' engineering and quality teams, most often at companies small enough that the algorithm team is one or two people.

Natural language processing

Clinical and technical documents: information extraction, span annotation and annotation-guideline design, de-identification, retrieval over internal corpora, and evaluation of language-model output where a wrong answer is expensive. The work is usually less about the model than about defining the label — which is where we start.

PICCdevice tip confirmed at the cavoatrial junctionfinding; no evidence of pneumothoraxnegated on post-procedure film. Continued on apixaban 5 mg BIDdrug.

Fig. 2 The note written after the procedure in Fig. 1. Negation is drawn dashed because a negated span is not an extraction — it records what was ruled out, and a system that misses it reads the note backwards. Synthetic text.

Method

In order. Each step is a gate, not a formality.

  1. 01

    Intake

    One conversation, a look at what data exists, and a question narrow enough to answer. No cost, no deck.

  2. 02

    Study

    Four to six weeks of hands-on work with weekly written notes, so nothing in the final report is a surprise.

  3. 03

    Report

    Specific, reproducible, and yours to keep — including the case against building, stated as plainly as the case for it.

  4. 04

    Build

    Only when the report supports it. Scoped from the study's findings rather than from an estimate made before the data was seen.

Fit

Said early to save both sides a month.

A good fit

  • Companies of roughly five to two hundred people
  • Data already collected, even if it is messy or unlabeled
  • Regulated products, where being wrong has a cost you can name
  • A technical person on your side who will read the report

Not a fit

  • General-purpose chatbots and assistant features
  • Strategy decks with no implementation behind them
  • Projects that need a vendor to say yes
  • Work that depends on data you do not have permission to use

Contact

Replies come from a person, usually within a day.

Tell us what data you have and what decision it needs to support.