Return to my storyAI Video Analytics

2024 - 2025

NeuralEye

A production-grade video intelligence platform for zero-shot brand and logo detection at scale.

Built as a multi-tenant AI system that combines zero-shot proposals, reference normalization, multi-modal verification, distributed inference, and observability to process long-form video without leaning on constant retraining.

Shipped a multi-modal layered detection system that improved end-to-end efficiency by 2x, supported concurrent long-duration jobs, and gave operators a usable surface for monitoring detections, quality, and tenant-safe analytics.

APIs

50+

across ingestion, detection, analytics, and reporting

Efficiency

2x

through layered inference, batching, and queueing

Architecture

Multi-modal

zero-shot proposals, reference prep, and verification

Architecture

System flow at a glance

The route from inputs to product surfaces, with the services and system layers that keep the workflow reliable in production.

Video Ingestion
Logo Reference Prep
Async Job Queue
Zero-Shot Proposal Layer
Multi-modal Verification
Embedding Matching
Analytics Storage
Ops Dashboard
Observability Stack

Problem

Brand and logo detection gets brittle when every new customer logo pushes the system toward retraining, manual tuning, or expensive review loops. NeuralEye needed a production path that could accept raw brand references, convert them into usable matching inputs, and run a layered zero-shot pipeline across long-form video while staying cost-aware.

Capabilities

AI SystemsZero-shot DetectionMulti-modal VerificationDistributed InferenceObservability

Key Decisions

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

Convert raw logos into cleaner comparable inputs so matching stays consistent across different client assets and video conditions.

Tradeoff: Reference quality becomes a product responsibility, not just a model concern.

Celery for distributed jobs

Enabled fault-tolerant asynchronous processing for high-volume workloads.

Tradeoff: Required stronger queue visibility and retry discipline.

Challenges

  • Handling varied brand assets, screen treatments, and noisy frames without retraining for every client.
  • Processing hours-long video without letting compute costs spiral.
  • Keeping tenant data isolated while still giving operators enough visibility to trust the pipeline.

Outcomes

  • Shipped a multi-modal layered detection flow that accepts brand references and routes them through proposal, matching, and verification stages.
  • Reduced diagnosis time by introducing structured observability around queues, detections, and job quality.
  • Delivered a React dashboard for monitoring pipelines, reviewing detections, and exporting tenant-safe analytics.

Stack

The parts of the stack doing the real work, from product surface to infrastructure and monitoring.

FastAPI

FastAPI

React

React

TypeScript

TypeScript

PostgreSQL

PostgreSQL

Redis

Redis

Celery

Celery

Docker

Docker

FS

FAISS

Prometheus

Prometheus

Grafana

Grafana

What's Next

  • Expand evaluation tooling around reference quality, segment confidence, and failure-mode review.
  • Add smarter orchestration for brand onboarding, replay workflows, and low-confidence human review.