Launching today
ChikitAI
Agentic AI for Healthcare Triage and Care Automation
37 followers
Agentic AI for Healthcare Triage and Care Automation
37 followers
ChikitAI is the healthcare agentic AI by NyuktAI that automates healthcare's messiest bottleneck: patient intake and triage, increasing patient intake capacity by 30% for healthcare providers. It talks to patients in natural language, captures a clinical-grade history, assesses urgency, and routes them to the right care. Built on proprietary clinical LLMs, it cuts wait times, reduces no-shows, and gives clinicians their time back.
ChikitAI
Refocus
The detail that stands out is routing on urgency rather than just symptom keywords. I build voice AI that checks in on aging parents by phone, and the toughest cases are the ones where someone underplays how they actually feel, so a confident triage decision can be exactly wrong. Does ChikitAI expose an uncertainty signal and hand borderline cases to a human, or does every intake get a definitive routing? Also curious how you validated the clinical LLM against real intake transcripts.
ChikitAI
@igorgurovich Thanks Igor. Voice AI is need of the hour. Great to know, you have built and tested it out. We have integrated external Voice AI agents for this. Happy to learn more about your Voice AI trained on clinical workflows.
On uncertainty: ChikitAI never issues a definitive routing. That was a deliberate architectural choice. Every assessment outputs a probability distribution across differentials with confidence scores. e.g. gastroenteritis 65% / appendicitis 20% / other 15%, not a verdict. That distribution goes to a clinician alongside the structured history, red flags, and the reasoning trace behind it. The doctor decides. We're a decision support layer at the first mile, not an autonomous router. In regulated care, a confident wrong answer is worse than a useful uncertain one, so we optimised for surfacing uncertainty rather than hiding it behind a single output.
Red-flag escalation runs as a separate rule-based path on top of the probabilistic layer. chest pain, breathing difficulty and similar trigger escalation regardless of what the model's confidence says. Probabilistic reasoning is good at the middle of the distribution and bad at the tail; deterministic rules cover the tail.
Genuinely curious how you're handling this on the voice side, voice gives you prosody and hesitation, which is signal we don't have in text. Does that actually help you catch the underplayers, or is it noisier than it looks?
Clinician here, and you're attacking the right end of the problem. The first mile is where the time actually leaks. My one worry is anchoring: once I'm handed a tidy structured history and a differential, it's tempting to reason from that framing instead of re-taking the story myself, which is exactly when I miss the thing the patient buried. In practice, do you see doctors re-interviewing, or mostly confirming the AI's version? Did you measure that shift at all?
ChikitAI
@hung_tran_from_notebook_os Thanks Hung. Great to have a feedback from a clinician. Much appreciated.
Anchoring is a real risk, and it's inherent to what we're doing. If you hand someone a tidy narrative, you've already shaped how they read the patient. That's the cost of the value we provide, not a bug we can engineer to zero.
What we do about it, in design:
We don't hand you a conclusion. Every assessment surfaces a distribution - 65% / 20% / 15% with confidence scores - not an answer. A ranked list of possibilities is harder to anchor on than a single tidy diagnosis. Deliberately less satisfying to read.
The reasoning trace is exposed, not hidden. You can see which symptoms drove which weighting. The point isn't to convince you, it's to give you something to disagree with. An opaque summary invites acceptance; a visible chain of reasoning invites challenge.
Structured chronology over narrative. We present symptoms across a timeline with severity rather than a smooth story. Stories anchor. Timelines with a gap in them make you ask about the gap.
Red flags escalate independently. Rule-based, running separately from the probabilistic layer, so the thing the patient buried and the model underweighted still surfaces on its own path.
Doctors using chikit do re-interview. At the begining it started with an objective to validate the AI's output, gradually it moved towards aiding them in focussing their questions on the narrower aspects rather than starting afresh.
If you'd be open to it, I'd genuinely like to talk more. Do try it out for yourself. Clinician input on what that study should look like is worth more to us right now than another feature. Either way, thank you for this one.
Node Health
Nice idea, always hate having to fill in a long form when entering a new clinic. Although when filling in these forms I usually share a lot of personal and health data. How do you ensure that my data remains completely private and is not used by whichever LLM you are using in the background? Do you store the data coming in on your own server as well, as it has to be passed on to the healthcare provider somehow?
ChikitAI
@renhuldt Thanks Ben. Just like, your physically filled details are kept securedly by Hospitals, our system provisions the same structure. The data shared by the patients are PII and are kept stored securely by the Hospitals within their infra. We don't send patient data to a third-party LLM. Our own clinical-grade LLMs are deployed as independent instances for the healthcare providers. Models train where the data already lives, inside the provider's environment. What travels back is model weights, not patient records. Every new deployment makes the model better without a single record leaving the institution. Nothing is aggregated into a central pool for training.
Would love to know, What's Node Health's approach to this? Curious whether you're seeing the same resistance to data centralisation, or whether it plays differently in your market.
One thing I'd love to see is multilingual support built in from the start, not as an afterthought. Healthcare serves incredibly diverse communities, and having intake conversations in a patient's preferred language right out of the box would make a real difference for accessibility and trust, especially for elderly or non-native English speakers who often struggle with current intake forms.
ChikitAI
@lknursafalge6s Thanks for the comment IIknur. You are spot on with identifying the language barrier at the very entry point of care. Forms, explaining the symptoms with their respective cultural context are major barriers for the physicians to get a full picture of the underlying case. Thats why Multilingual isn't a layer we added on top. It's where we started. The reason is that most systems treat language as a translation problem. Patient speaks, text gets converted to English, then the clinical model reasons over the English. But meaning leaks at that seam. People describe pain, duration, and severity through idiom that doesn't survive translation. By the time the model sees it, the nuance is already gone, and the model is reasoning confidently over a degraded input. So our clinical LLMs convert the patient's own language directly into medical terminology, no English hop in the middle. The reasoning happens on what they actually said. Voice input is part of this too, for anyone who finds forms hard for reasons that have nothing to do with language.
Curious which languages or communities you had in mind? Genuinely useful input for how we prioritise.