What the Karpathy Loop is — and why it changes medical-device sales
Andrej Karpathy described AI training as a closed feedback loop where every example improves the model. Solara applies the same architecture to a sales network: every rep interaction trains the system that targets the next rep interaction. Here is how it works, and why it compounds.
Most field sales organizations treat the rep visit as a one-way push. The rep walks in with a script, makes a pitch, and walks out. Whatever happens next — adoption, stall, churn — feeds back into a CRM as a status flag, and that’s where it stops. The next rep’s pitch doesn’t change. The targeting model doesn’t update. The organization gets older but not smarter.
That’s the gap the Karpathy Loop closes.
Where the term comes from
Andrej Karpathy, in his work on neural networks, described AI training as a closed loop: the model produces an output, the output is evaluated against ground truth, and the error signal flows back into the next iteration of the model’s weights. The loop runs continuously, and every example seen by the system makes the next prediction sharper. The model isn’t a static asset — it’s a process.
Most software organizations treat their sales process as a static asset. Solara treats it as a process.
What the loop looks like in a sales network
Four components, strictly separated:
- The mutating file. The thing that changes between iterations: rep outreach scripts, email templates, call openers, objection handlers. The platform generates variants and deploys them to reps.
- The locked eval. Real-world physician conversion data — adopt, trial, stall, no-go. The eval is never edited by the system. It comes from the field, attached to the script variant the rep used.
- The direction. Human-written goals, segments, compliance constraints. Rules of the game. Decided by people, fed into the loop as a stable input.
- The reviewer. A separate evaluator (in our case, a model-driven one with human approval gates) that scores variants against direction and eval data, flags compliance issues, and recommends which variants to retire and which to scale.
Why the separation matters
This is the part most teams skip and the part that determines whether the loop works.
If the same agent that writes the scripts also evaluates them, the system over-rates its own work. If the eval data can be edited by the agent during a cycle, the eval drifts toward whatever the agent already wants to believe. If direction is implicit, the loop optimizes for whatever it can measure, not for what the business actually wants.
The discipline isn’t a UX preference. It’s the structural property that makes the loop trustworthy at scale.
What changes for the manufacturer
Three things, measurably:
- Targeting tightens. Physician-level outcome data flows into the next iteration of the targeting model. Reps walk into shorter lists of higher-likelihood adopters. CAC drops without rep headcount changing.
- Scripts compound. A variant that converted well in Texas wound clinics gets surfaced for Florida wound clinics in the next cycle. The best rep in the network effectively trains the others. The floor rises every quarter.
- Stall signals show up earlier. Adoption-state tracking — contacted, met, trialed, adopted, churned — runs in real time. Manufacturers see the problem in week six, not at end of quarter when the dashboard finally updates.
What the rep experiences
From the rep’s seat, the loop looks like a list. They open the platform in the morning and see a ranked list of physicians most likely to adopt this week, paired with the script variants that have been converting in their segment. They run the calls, log outcomes, and the list updates overnight.
The rep isn’t making the targeting choices anymore — the system is — but the system is trained on the entire network’s outcomes, not on the rep’s memory of which doctors return their calls.
Why this is hard for incumbents to copy
The Karpathy Loop only works when you have closed-loop telemetry. Most legacy distribution organizations don’t. Their conversion data lives in the manufacturer’s claims feed; their script variants live in a folder; their direction lives in a quarterly off-site whiteboard photo. The infrastructure to close the loop is the moat.
Solara was built loop-first. Every rep interaction creates the data that trains the system that targets the next rep interaction. Year-two reps walk in with a tighter conversation than year-one reps did. New manufacturers joining the network inherit the floor that prior manufacturers helped raise. The network doesn’t just scale — it compounds.
Where to go from here
If you’re a manufacturer evaluating distribution, the question to ask any potential partner is: show me your loop. Show me the eval. Show me the variant log. Show me the reviewer. If they can’t, the system isn’t learning — it’s just running.