entering the universe
v
Videet
now · behaviour cloning in PufferDrive

Hi,I’mVideet.

vi-DEETBangalore, India

I build computer vision systems for cars at Valeo — LIDAR perception for ADAS, two patents in flight, and a network running at 13.5 ms on TI hardware.

On the side I’m building clarzo.ai to help Indians make sense of their money, and listenai.in — voice AI for humans, not call centers.

say hi→
or see what I’m working on↓
scroll
02 — origins

A journey across
curiosity & code.

Six years of training models, shipping products, and writing about the small details that make perception feel almost human.

201601

Joined IIT Kharagpur

Dual degree in Industrial Systems Engineering and Management. The years that turned curiosity into discipline.

202102

Founder of listenai.in

Co-founded listenai.in out of college — a conversational AI platform for behavior, emotion, and intent across voice, text, and video.

202103

Joined AgNext

Moved into food-AI as a Machine Learning Engineer — CNN pipelines for grain quality assessment, image-based sample weighing, and an Airflow + MLflow training stack.

202304

Joined Valeo

Began leading the LIDAR perception system for ADAS. Architected the VLPS network — a patented architecture running at 13.5ms on TI hardware. Currently here.

202605

Founded clarzo.ai

Parallel to Valeo, founded clarzo.ai — an AI wealth-intelligence platform helping Indian investors see their money clearly.

0+
Years in ML
0
Patents
0
Companies founded
03 — companies

Two companies,
one philosophy.

Build small, ship slowly, respect the user's attention. One product is live and helping Indians make sense of their money. The other is quietly being built around a different question — what would voice AI look like if it weren't designed for call centers?

2026Live·clarzo.ai

Clarzo

An AI wealth-intelligence platform helping Indian investors finally see their money clearly.

what it does

Clarzo unifies stocks, mutual funds, EPF, FDs, SIPs, and crypto into a single, honest view of your money. You can talk to it the way you'd talk to a financial colleague — ask where you're overexposed, what to do with a bonus, whether you're still on track for the goal you set two years ago. When AI isn't enough, SEBI-registered advisors are built one tap away.

why you'll use it
  • →Unified dashboard across 10+ asset classes — no more juggling eight apps and a spreadsheet
  • →Conversational AI fluent in Indian retail finance — not a generic chatbot, a colleague
  • →SEBI-registered advisors routed inside the product when you want a human in the loop
how your life shifts
  • ◆You stop opening five apps just to figure out where you actually stand
  • ◆Bonuses, SIPs, and rebalancing become confident decisions — backed by your real numbers
  • ◆Money stops being a low-grade anxiety running in the background of your life
Open Clarzo↗
2021Live·listenai.in

listenai

A voice-first AI for the small, human moments — built for consumers, not call centers.

what it does

Most voice AI today is built to make enterprises more efficient — sales calls, transcripts, call-center summaries. listenai.in is the opposite: a voice product designed for everyday life. A two-minute reflection before bed. A quiet thought you'd rather say out loud than type. A daily check-in with something that listens without judging. Calm by design, useful by default.

why you'll use it
  • →Voice that listens without interrupting — no AI-host energy, no performative warmth
  • →Built for moments, not minutes — no streaks, no engagement loops, no metrics chasing you
  • →Privacy first — your voice is yours, by design and by architecture
how your life shifts
  • ◆Replace twenty minutes of doomscrolling with a two-minute voice journal
  • ◆Talk through a problem with something that won't judge, push, or sell
  • ◆A daily quiet check-in that's yours, private, and kind
Open listenai↗
04 — computer vision lab

Where pixels
become understanding.

A small, ongoing collection of research and production work in computer vision — segmentation, tracking, restoration, and the quiet places between.

pipeline · architecture
Sensor
Backbone
Neck
Head
Output
FeaturedADAS
live · simulating 1,000+ agents
ego online
2025 — NowResearcher · Contributor

PufferDrive

A fast, ADAS-focused driving simulator for stress-testing driver-assistance behaviors against real-world traffic at the scale of the Waymo Open Motion Dataset.

100K
Scenes
Py · C · CUDA
Languages
77+
Stars
PufferLibC / CUDAPythonRaylibWOMDEmscriptenADAS Validation
Read case study→RepositoryLive demo →
Researchopen →
Input
Predicted

Low-Light Semantic Segmentation

Real-time scene parsing under 0.1 lux conditions.

78.4
mIoU
62
FPS
12M
Params
Productionopen →
Input
Predicted

Long-Range Object Tracker

Persistent tracking across 30s+ occlusions in dense scenes.

84.1
MOTA
−63%
ID switches
11ms
Latency
Researchopen →
Input
Predicted

Diffusion Image Restoration

One-step diffusion model for blind image restoration.

32.7
PSNR
1
Steps
180M
Params
05 — journal

Field notes
from the lab.

Research2026-04-12 · 7 min

The Quiet Craft of Attention

What we lose when our models stop noticing the small things.

Read essay→
"What we lose when our models stop noticing the small things."
Startups5 min

How to Build Small

Notes on running a tiny studio without losing your mind.

2026-03-08→
AI9 min

What the Camera Sees

A meditation on perception, and the long road from pixels to meaning.

2026-02-02→
06 — notes

A working
notebook.

ML topics I keep re-explaining to myself — architectures, training mechanics, LLMs, vision, MLOps. Some pages are written, most are stubs I’m filling in over time.

144topics across 13 categories
1 written·143 drafts
open the notebook→
a slice of the index
  • Transformers
  • Attention
  • Diffusion Models
  • Reinforcement Learning
  • RAG
  • FlashAttention
  • BERT
  • GPT
  • Vision Transformer (ViT)
  • Bias-Variance Tradeoff
  • + 134 more
07 — resume

The long-form
version of me.

Download PDF
experience

Leading the LIDAR perception system for ADAS. Architected VLPS (Valeo Lidar Point-Cloud Segmentation, patent 2024PF00332) — running at 13.5ms on TI hardware with 90% precision after a 30% prune. Authored a second filing on a hybrid neural network for LIDAR point-cloud segmentation in ADAS applications (patent 2024PF02608). Built an end-to-end CI/CD pipeline on GCP Vertex AI for multitask 2D/3D point-cloud training, with GSAM-based preannotation of raw point clouds.

Built CNN models for digital quality assessment of grains, oilseeds, and spices — segmentation, classification, and object detection across identification, quantification, and feature localization. Developed image-based sample-weight estimation that removed physical weighing from the pipeline. Owned the Airflow + MLflow stack for data acquisition, training, and deployment.

Founding member of a conversational AI platform for behavior, emotion, and intent across voice, text, and video. Built voice clustering with HMMs, neural nets, and Transformers; led a team of five to ship the initial product and reach 54% MoM growth across three early partners in the first four months.

B.Tech + M.Tech. Master's thesis on visual saliency — SALGAN and SALNET trained on SALICON, paired with eye-tracker data captured while driving. Bachelor's thesis on a sensor-embedded IoT safety jacket for mine workers — GPS, pulse, temperature, gas, and a real-time alert system.

skills
Computer Vision
96
PyTorch / TensorFlow / Keras
92
Python / C++
90
MLOps · Vertex AI, MLflow, Airflow
86
Edge & On-Device Inference
82
Natural Language Processing
78
patents
Valeo Lidar Pointcloud Segmentation (VLPS)
Patent · 2024PF00332 · 2024