APIs
50+
REST APIs delivered across the product
Opening the field notes · Systems
Unfolding the system story.
Skip to the storyKruthik GowdaSenior Software EngineerContact 
Back to Request Race BenchAI Video Analytics
selected work
NeuralEye
Find and inspect brand appearances in long-form video.
NeuralEye finds possible brand appearances in long-form video and lets people inspect each proposed match against its source frame, timestamp, bounding box, and confidence score. Under the hood, it matches SigLIP embeddings for each logo reference against a tenant-scoped Qdrant vector index.
My role
Ketrium Labs · co-founder and engineer
Contribution and scope
I architected multi-tenant brand detection for long-form video and built the distributed workflows, API surface, and monitoring tools.
Execution model
Parallel
Inference used batching and caching for long-form video.
Detection flow
Layered
zero-shot proposals, reference prep, and verification
Capability map
Product capabilities at a glance
Related capabilities, grouped for exploration. This is a capability index, not a diagram of internal services or execution order.
Video Ingestion
Logo Reference Prep
Zero-Shot Proposal Layer
Selective Verification
SigLIP / Qdrant Matching
Analytics Storage
Ops Dashboard
Observability Stack
Separate from the production architecture
Related teaching model
These independent experiments explore skills connected to this work. They are not representations of specific production components.
Problem
NeuralEye needed to compare new brand references with long-form video, keep proposed matches open to inspection, and manage compute cost.
Capabilities
AI SystemsZero-shot DetectionSelective VerificationDistributed InferenceObservability
Design considerations
Tradeoffs suggested by the documented capabilities.
Layered zero-shot inference
Use a proposal, matching, and verification flow so the system can adapt to new brand assets without retraining the whole stack.
Tradeoff: Adds coordination, confidence calibration, and evaluation work across more moving parts.
Reference normalization before matching
SigLIP embeddings are matched against a tenant-scoped Qdrant vector index. Normalizing raw logo references helps keep comparisons consistent across different client assets and video conditions.
Tradeoff: Reference quality becomes a product responsibility, not just a model concern.
Route async work by workload
Distributed GPU tasks run through Celery; complex workflows are coordinated through Temporal.
Tradeoff: Two execution paths make it important to follow task state through the right worker or workflow.
Challenges
- Handling varied brand assets, screen treatments, and noisy frames without training a detector for every new brand.
- Processing long-form video efficiently while managing compute cost.
- Keeping tenant data isolated while still giving operators enough visibility to trust the pipeline.
What shipped
- Shipped a layered zero-shot detection flow that accepts brand references and routes them through proposal, matching, and selective verification stages.
- Added monitoring for distributed workflows and detection quality.
- Delivered a React dashboard for monitoring pipelines, reviewing detections, and exporting tenant-safe analytics.
Named technologies
Tools used in this system.
FastAPI
React
TypeScript
PostgreSQL
Redis
Celery
T
Temporal
Docker
S
SigLIP
Q
Qdrant
C
ClickHouse
Prometheus
Grafana
Further exploration
- Explore evaluation tools for reference quality, segment confidence, and failure review.
- Explore orchestration for brand onboarding, replay workflows, and human review of low-confidence detections.