Shubham
Jakhmola
Building agentic AI systems that hold up in production.
Conversational agents, retrieval, and the evaluation that keeps them honest. Based in Delhi.
buster · shubham's agent · v1
Hi, I'm Buster. Ask me anything about Shubham's work.
“{{ agentAsked }}”
{{ s.icon }}
{{ s.label }}
{{ s.time }}
“{{ agentAsked }}”
{{ agentReply }}
{{ mascotHint }}
{{ mascotHint }}
{{ mascotHint }}
{{ autoLabel }}
{{ t }}
{{ sel.cityName }}{{ sel.remoteTag }}
{{ sel.years }}
{{ sel.role }}
{{ sel.org }}
{{ sel.metric }}
{{ sel.metricLabel }}
→{{ pt }}
{{ cs.tag }}
{{ cs.year }}
{{ cs.title }}
{{ n.label }}
{{ n.sub }}
{{ nodeTip.label }}
{{ nodeTip.text }}
01
Capture: six sensors, every session
20 users × 15 sessions · raw events, no labels beyond who held the phone
xy
accelerometer
gyroscope
magnetometer
gravity
rotation
user accel
02
Describe: one session becomes one vector
198statistical
162temporal
126frequency
78magnitude
~45x-sensor
~609 →200 kept · mutual information
03
Decide: who is it, and how sure?
Random Forest · 94.7%
> 80%28.1%
allowauto-pass
50–80%54.2%
soft verifypush notification
< 50%17.7%
hard verify2FA / PIN
01
Stack: three MRI modalities, one volume
FLAIRT2T1GdT1 dropped
→crop128³ around the tumour centre
rescaleMinMax → 0–255
correctbias field
02
Learn: a 3D U-Net, against two backbones
encode 16 → 256decode + skip links
3D U-Netchosen
U-Net + ResNetencoder
U-Net + VGGencoder
loss: dice + focal · Adam
03
Segment: every voxel gets a class
test slice
→prediction
=
overlay
necrotic core
edema
enhancing
0.73 mean IoU · test
{{ sp.t }}
{{ sp.s }}
write to me
j4khmola@gmail.com
Open to full-time roles and contract work on agents, RAG and evaluation. Based in Delhi, happy to work remote.
© 2026 · drawn in light · drone model by LaVADraGoN · CC BY 4.0
buster · shubham's agent · v1
{{ contactLine }}
“{{ agentAsked }}”
{{ s.icon }}
{{ s.label }}
{{ s.time }}
“{{ agentAsked }}”
{{ agentReply }}