Kishore Nuthalapati
Founder · Researcher

Kishore Nuthalapati

Building AI systems that know what they don’t know.

I co-founded Aptivae, a hiring verification company, and I founded Second Thought Initiative, a nonprofit that teaches high schoolers AI literacy. I do research on AI calibration, and I’m building Steady Arm, a prosthetics project. I’ve always been a builder.

Currently
Aptivae is fully working and going to market
Raising a $250K pre-seed
In talks with VC funds, accelerators and incubators
Steady Arm research paper in preparation for submission

At a glance

Students reached
10K
Aptivae beta users
50+
Research acceptances
8 7 ICML + NeurIPS
Aptivae funding
Received
Calibration error
−74%
Aptivae pre-seed round
$250K target, open
Accuracy gain
+14.7 pts
Steady Arm raised
$1.5K

Work

What connects it

Most of it comes back to one question: how far off is the system, and does it know? I studied it in AI reasoning, I’m applying it to hiring, I’m raising to apply it to prosthetics, and I teach it to high schoolers.

Co-founder

Aptivae

Aptivae is an AI-native verification layer for hiring. It finds a candidate’s exact skill gap, closes it with a targeted sprint, and proves readiness on graded work, so an employer doesn’t have to gamble on a resume.

At a glance

Beta users
50+
Pre-seed round
Open $250K target
Product
Fully working
Funding
Received
Candidate price
$0
Employer price
12% per verified hire

Where it is

  • MVP built
    Done
  • 50+ people tested the MVP
    Done
  • Fully working product
    Done
  • Going to market through our go-to-market team and CMO
    In progress

How it works

  1. 01
    Apply
    Submit the job, the full description and a resume.
  2. 02
    Assess
    Find the exact skill gap and a readiness score.
  3. 03
    Sprint
    A short, targeted sprint on what’s missing.
  4. 04
    Graded project
    Real work, graded against a rubric built by a domain expert.
  5. 05
    Certified
    Proof the candidate holds and shares.

Every applicant gets one of three answers

  • Apply Now
    Verified and ready for the role today.
  • Train Then Apply
    Close, with a defined gap to close first.
  • Not Yet
    Not a fit right now, and told plainly.

What it does, and what it won’t

Grades
Real, submitted work against a rubric built by a human domain expert.
Verifies
The specific requirements of the role, not a generic average.
Shows
Employers graded work before they meet anyone, and candidates written feedback instead of a silent rejection.
Issues
A certificate the candidate holds and shares themselves.
Won’t
Auto-apply on a candidate’s behalf, mass-submit resumes, or guess fit from resume keywords. A verdict is evidence, and a person still decides what to do with it.

Business model and my role

Candidates
Free.
Employers
12% of first-year salary, per verified hire. No subscription.
My role
AI research, investor relations, fundraising, company build-out.
Go-to-market
A go-to-market team and a CMO, getting the product to more people.
Fundraising
Funding received. Raising a $250K pre-seed, round open. In talks with VC funds, accelerators and incubators.
Builder

Steady Arm

Steady Arm is a personal prosthetics project. I built it myself, I’ve raised $1.5K for it so far, and I’m raising more to turn it into a system that reduces error across prosthetics instead of in one device.

At a glance

Raised so far
$1.5K
Status
Personal project
People tested, offline
91
Research paper
In preparation

Research paper

Title
Misfire-Rate Guarantees for Myoelectric Prosthesis Control: Conformal and Adaptive Rejection Thresholds
Status
In preparation for submission to venues.
The problem
Prosthetic hands usually act only when a classifier’s confidence passes a fixed number, often 0.90. That number says nothing about how often the hand will move the wrong way for a new user, or after conditions change.
The idea
Set the threshold from a target instead, such as “misfire on at most 5% of decisions,” and keep that promise for each user.

Results

People tested
91
Public datasets
4
Default threshold after a 45° armband shift
46.1% misfires
Adaptive threshold, same test
4.9% mean
Worst user, adaptive (5% target)
5.0%
Hand acts, adaptive vs best fixed
31.5% vs 14.0%
New arm positions, default vs adaptive
20.5% to 4.0%
Guarantee held
1,365 of 1,365 runs

Thresholds tuned on training data exceeded every target tested with shared models. Conformal thresholds met the target on average at every level.

Limits and links

What this is
Offline tests on recorded muscle signals from public datasets. Able-bodied participants, no physical hand yet.
Next
Multi-day and amputee data, implicit feedback, and real-time evaluation.
Code
SteadyHand on GitHub (open source, built on LibEMG)

Today and next

Today
A prosthetic arm in active development, plus the control software that decides when it moves. The paper tests that software offline on recorded muscle signals, without a physical hand.
Next
A system that aims to reduce error across prosthetics, measured by how much error it removes.
Why this one
Same question as my AI research: how far off is the system, and can the error be shrunk? In prosthetics the error is physical.
Researcher

Research

My research is about AI calibration: the gap between how confident an AI system sounds and how often it’s actually right, and how to close it. Most of it is on reasoning and clinical dialogue.

At a glance

Acceptances
8
Calibration error
−74%
ICML 2026 workshops
7
Accuracy gain
+14.7 pts
NeurIPS 2026
1
Papers in progress
2

What calibration means

predicted confidence → actual accuracy →
ECE is Expected Calibration Error: the average gap between a model’s stated confidence and how often it’s correct. Points on the dashed line are perfectly calibrated. The two orange points are overconfident. Lower ECE is better. This is an illustration, not data from the papers.

Papers

−74%
calibration error, +14.7 pts accuracy

Probabilistic Chain-of-Thought: Sequential Bayesian Inference over Latent Reasoning Correctness

with Suriya Dev Saravanakumar, Ezra Wesenie and Laksh Patel
Accepted at 7 ICML 2026 workshops
AI4MathCTBCoLoRAICompLearnEIMLSCALESPIGM

Cuts Expected Calibration Error by 74% and improves accuracy by 14.7 percentage points on MATH and GSM8K.

3 × 2
LLMs × reasoning modes

Calibration Collapse in Sequential Clinical Dialogue Agents

Accepted at NeurIPS 2026

How AI confidence degrades across multi-turn clinical conversations, comparing GPT‑4.1, gpt‑oss‑120b and DeepSeek‑v4‑flash in thinking and non-thinking modes.

4.9%
misfire after a 45° shift, vs 46.1% for the default

Misfire-Rate Guarantees for Myoelectric Prosthesis Control: Conformal and Adaptive Rejection Thresholds

In preparation for submission

Chooses a prosthetic hand’s confidence threshold from a target misfire rate and keeps that promise for new users and after conditions change. Tested on 91 people across four public datasets; the guarantee held in all 1,365 runs. More on Steady Arm

Cureus
target journal

The Broad-Footprint Hypothesis

In progress

A narrative review proposing that arousal-induced neuromodulation during trauma makes the brain’s short-term consolidation system operate more broadly than usual, which would explain why PTSD memories feel so total and resist extinction therapy. Advised by a PhD candidate through the Lumiere Individual Research Program.

Founder · Nonprofit

Second Thought Initiative

A nonprofit that teaches high schoolers AI literacy, built on one skill: knowing when to trust an AI’s output and when to push back on it.

At a glance

Students reached
10K
Kits
Distributed to students

The program

Who
High schoolers, including students at Title I schools.
The one skill
Telling an AI that’s actually right from one that just sounds right.
Reach
10K students, with kits distributed.
Grounding
My own calibration research. See the research
← back

on enough

greed

i keep trying to find the line between ambition and greed and i keep not finding it.

the dictionary version is easy. greed is wanting more than you need. but "need" is doing a lot of work in that sentence. need for what? to eat, to be safe, to not feel behind? nobody i know chases a number so they can eat. the need is always something else standing behind the number, wearing it like a costume.

sometimes i pick a number i'd be fine with. really fine, no asterisk. then i picture hitting it, and i don't stop. i can feel that i don't stop, even in the daydream. which tells me the number was never what i wanted. it was standing in for a feeling, and the feeling doesn't live in the number, so i go looking for it a few zeros up.

i think that's the part of greed i can't argue with. you put the feeling somewhere out ahead of you and walk toward it, and it moves at exactly your walking speed. i doubt that's an accident.

but then i get stuck on the other side. a lot of good things exist because somebody wanted more than was reasonable. nobody starts a company, or spends two years on a study, out of a calm sense that things are fine as they are. dissatisfaction is a fuel. an ugly one, but it burns. if i could switch my wanting off i'm honestly not sure i'd do any of the things i'm proud of.

so what separates the two? the best i've got is this. ambition points at a thing, and greed points at a direction. you can arrive at a thing. you can't arrive at more.

i don't know if that's a real distinction or just a nicer way of describing my own wanting. it does give me one question i can use, which is whether i can say what done looks like. when i can't, i get suspicious of myself.

the other tell is comparison. i've caught myself wanting something mostly because someone else had it. that one i'm pretty sure is greed in its purest form, because there's no feeling behind it at all, just a gap between two numbers.

i still don't know where enough is. i know it isn't a number, because i've tried that. maybe it's closer to a sentence you can say out loud and mean, like "i'd be okay if it stopped here." i can say it on some days. i'd like it to be more days, without it happening because i stopped wanting things, since then i'm back to not building anything.

← back

on being right by accident

success

most of what i study is the gap between being right and knowing why you're right. a model can land on the correct answer and still be reasoning badly, and from the outside you can't tell. the answer looks the same. it only shows up when you run it a hundred times on slightly different problems and the luck stops covering for it.

i've started to think success works the same way, and that it's a bigger problem than the one in my papers.

when something works, it comes with a story attached. you did x, then y, then it worked, so x and y were why. nobody hands you the other timeline, the one where you did the same x and y and nothing happened, and that's the only way to know whether x and y mattered. you get one run. you win, and your brain, which is a very confident system, files the result under "what works."

this is why successful people give such bad advice, and i say that as someone who'd like to be one. they aren't lying. they're reporting, sincerely, the story their one run produced. the confidence is real. the calibration isn't.

failure gets the opposite treatment, which is almost funny. when something goes badly you comb through every decision looking for the mistake, even when the decision was fine and the dice were just mean. so the same quality of thinking gets graded two different ways depending on how it landed.

i don't have a clean way out. a decision journal, maybe, where you write down what you expect before you find out. i've done it in bursts. it's humbling in a way i can't fully explain, because you learn how often you wrote "this will definitely work" next to things that were a coin flip.

what i can't resolve is what to do with it. if success is partly luck and you can't tell how much, do you take the credit? take less? taking too little feels like its own dishonesty, because there were real choices and some of them were good. taking all of it makes you stupid slowly, because you stop updating, and the person who thinks they've figured it out is the one most likely to run the same play after the board has changed.

so i'm trying to carry two numbers. one for how much of this was me, and one for how sure i am about the first. the second is usually lower than i'd like.

maybe that's most of growing up. the amount you know doesn't go up much. the estimate of what you don't know gets more accurate. i'm not there yet.

← back

on being content

contentment

there's a version of contentment i'm afraid of and a version i want, and from the outside they look the same.

the one i'm afraid of is comfort that showed up early. nothing hurts enough to move you, the temperature is fine, and ten years go by at that temperature. you can watch it happen to people who were sharp at eighteen. nobody decides it. it's just a lot of days where staying was slightly easier than leaving.

the one i want is quieter. it's being okay with where you are while still being bothered by things that are actually wrong. i think those two can coexist, but it takes some effort, because being at peace with yourself and being at peace with a broken thing are different, and the second one is just giving up with good posture.

my problem is that i can only hold the good version for about forty minutes. i'll finish something and feel good, a clean quiet kind of good, and then i open my phone and within a few swipes someone my age has done something that makes mine look small. i'm not sure it's even envy. my sense of enough gets overwritten by someone else's, and theirs is always bigger, because the feed only keeps the biggest ones.

the feed is a strange object when you think about it. it's a list of everyone's best day, sorted by how much it'll make you feel something. nobody in it is lying. every item is true and the total is still wrong, the way a city looks wrong if you only ever see its postcards.

i've tried a few fixes. deleting the apps works for about a week. reminding myself it's curated works in my head and not at all in my stomach. the only thing that has ever worked is being in the middle of something hard enough that i forget to compare. when i'm deep in a problem i'm content in a way that has nothing to do with how i'm doing, because i'm not keeping score. which makes me wonder if contentment is even a place you can reach. maybe it's a side effect of paying attention, and you only get it when you aren't looking for it.

that's annoying, because it means you can't go and get it. you can only arrange your days so it happens more often and then leave it alone.

i haven't figured out how that fits with wanting to build big things. you can't be fully inside the work and also thinking about where it lands. maybe the answer is to take turns: look up at the scoreboard on sundays, head down the other six days. i'm not sure that counts as a philosophy. it might just be a schedule.

← back

on choosing

hiring, aptivae

every hiring decision is a bet placed with bad information, and everyone involved knows it and acts like they don't.

a resume is a list of things someone says about themselves, written to get picked. an interview is forty minutes of two people performing confidence at each other. some of the people who get chosen are great. some of the great ones get cut in the first thirty seconds by a filter that never explains itself. from the outside it all looks like a decision. mostly it's a guess that got a title.

the guessing doesn't bother me as much as what it does to the people the guess went against. most rejections come with nothing. no reason, no gap named, no next step. you get a no, or more often you get silence, and you're left to work out whether the problem was your skills, your resume, your timing, or you. that's a lot of information to hold back from someone, and it's the exact information that would let them improve.

this is where aptivae came from. i co-founded it on a fairly simple belief: if you can verify what someone can actually do, you don't have to guess, and if you don't have to guess, you can be honest about the gap. so there are three answers instead of two. apply now, for people who've shown they can do the job. train then apply, for people who are close and can see exactly what's missing. and not yet, which is a different thing from no. no ends the conversation. not yet tells you where you are.

i think about that third answer more than the other two. it's uncomfortable to say to someone, and most systems avoid it by saying nothing, which feels more polite and is actually crueler.

the evidence so far is small and i'd rather say so than dress it up. more than fifty people have tested the mvp, and the product now works end to end, which is a different feeling from a prototype that only works when you're standing next to it. what i don't know yet is whether it holds up with people who've never heard of us. that's the next stretch: getting it in front of strangers, with a go-to-market team and a cmo handling the part that has nothing to do with building it. fifty testers can tell you a thing isn't broken. they can't tell you it's wanted.

i go back and forth on whether verification is even the right frame. it assumes the thing worth knowing about a person can be measured, and some of it probably can't. a test won't tell you who's good to have next to you at 11pm when a launch is breaking. but we're currently choosing people on a worse proxy than that, and "can this person do the work" should get an honest answer before anyone's vibe does. i'm not sure that's completely right. it's what i'm building on until i find out.