Karthik Iyer

Computer Vision Engineer·System Builder / Architect

Bangalore, Karnataka · Building in public since Mar 2020 (~6.3 yrs)

500+Inspection sites deployed

94%Damage-type classification accuracy

4xInference speedup after optimization

30 FPSSustained frame rate on edge

Growth trajectory

From detection models to production vision systems.

A clear shift from shipping detection models to owning full vision pipelines - deployment, optimization and reliability included.

Mar 2020Shipped first detection model

Aug 2021Solved production edge-case failures

Feb 2022Built real-time tracking

Nov 2022Moved into InsurTech vision

Aug 2024Owns edge deployment optimization

Mar 2025Exploring vision-language models

Mar 2020
Shipped first detection model
Aug 2021
Solved production edge-case failures
Feb 2022
Built real-time tracking
Nov 2022
Moved into InsurTech vision
Aug 2024
Owns edge deployment optimization
Mar 2025
Exploring vision-language models

Signals that matter

Live in Production.

Each signal ties back to a specific, evidenced moment - not a self-rating.

Built2

Vision systems shipped end-to-end, from model to deployed pipeline.

EvidenceBuilt an end-to-end vehicle damage assessment pipeline

Scaled2

Taken from a working model to real inspection volume in production.

EvidenceDeployed real-time multi-object tracking across 500+ inspection sites

Solved1

Detection failures and edge cases closed under real-world conditions.

EvidenceSolved false-positive detections on cluttered shelf scenes

Led1

Design and build decisions for a vision pipeline owned solo.

EvidenceLed the OCR pipeline build for vehicle documents

Stack

Daily tools.

What Karthik actually reaches for, pulled from the tags attached to captured moments - not a keyword list.

Work

Key Projects By Role

Case studies, not bullet points - impact numbers attached to each one.

ClaimSight AI

Senior Computer Vision Engineer · Nov 2022 - Present

5 months

Built an end-to-end vehicle damage assessment pipeline

Architected object detection and segmentation models that classify vehicle damage type and severity directly from claim photos and video, cutting manual inspection out of the fast path entirely.

94% damage-type accuracyclaims turnaround cut by 3 days

Object DetectionSegmentationPyTorch

Self-reportedView case study

4 months

Deployed real-time multi-object tracking for video claims

Built a tracking pipeline for video-based claims verification, running on-device at inspection sites rather than round-tripping video to the cloud.

30 FPS on edge devicesdeployed across 500+ sites

Multi-Object TrackingVideo Analytics

Verified via GitHubView case study

3 months

Optimized inference with TensorRT + quantization

Took the detection and tracking models through quantization and TensorRT compilation to fit within edge-device memory and latency budgets.

4x inference speedup70% smaller memory footprint

TensorRTQuantization

Verified via MediumView case study

2 months

Led the OCR pipeline build for vehicle documents

Owned the design and build of an OCR pipeline extracting structured fields from registration and insurance documents photographed in the field.

97% field-extraction accuracy

OCRImage Processing

Verified via Public ProfileView case study

2 months

Piloted Vision-Language Models for auto damage descriptions

Explored VLMs for generating a plain-language damage description alongside the structured classification, for the human adjuster to quickly sanity-check.

Piloted on 2 claim types

VLMsVision Transformers

Self-reportedView case study

Vion Robotics

Computer Vision Engineer · Mar 2020 - Oct 2022Verified · employer record

4 months

Built an object detection pipeline for retail shelf monitoring

Built and trained detection models identifying out-of-stock and misplaced items from in-store camera feeds, the team's first production CV deployment.

First production CV system at the org

Object DetectionOpenCV

Verified · employer recordView case study

6 weeks

Solved false-positive detections on cluttered shelf scenes

Traced a spike in false out-of-stock alerts to poor generalization on visually cluttered shelves, and closed the gap with harder negative mining during training.

45% fewer false positives

Model DebuggingData Augmentation

Verified · employer recordView case study

Outside work

What gets built off the clock.

Experiments, writing and talks - separate from paid work, tagged as such.

Open source

trackeval-lite

A lightweight multi-object tracking evaluation toolkit - the harness Karthik wished existed while benchmarking tracker swaps.

Verified · GitHub★ 95

Article

Why your detector's mAP doesn't predict production performance

On the gap between benchmark metrics and what actually breaks once a model meets real camera hardware and lighting.

Published Jan 20256 min read

Speaker

Deploying Vision Models at the Edge

Talk on the trade-offs between accuracy, latency and memory when moving vision models off the cloud.

Bangalore CV Meetup · Feb 2025

Experiment

Browser-based damage detector demo

A proof-of-concept running a quantized detection model client-side via ONNX.js, no server round-trip required.

Prototype · unshipped

Publication / book

Nothing published yet - this will fill in the moment it happens.

Learning

Formal record.

B.E., Computer Science
RV College of Engineering, Bangalore
2016 - 2020
NVIDIA Deep Learning Institute - Computer Vision
Verified · issuer
2023
Vision Transformers and Multimodal Models
Independent study
2024

Captured moments

Not a work log. A record of momentum.

Each shaded day is when Karthik captured something built, solved, scaled or led - density here reflects real activity, not tenure.

LessMore

Open to

Open to the right opportunity.

Roles, speaking, mentoring, or a good conversation - Karthik is listening.

Open to rolesOpen to speakingOpen to mentoringOpen to relocationOpen to advising

Connect with Karthik