PESU Venture Labs

Overview

PESU Venture Labs is a university-affiliated technology incubator where student developers work on real client projects. Unlike traditional internships or academic projects, we delivered production-ready solutions to paying external clients—real businesses with real problems and tight deadlines.

Over two years as an Applied AI Developer, I worked across diverse domains: computer vision, semantic search, generative AI, and document processing. Each project taught me valuable lessons about bridging the gap between research and production.

Major Projects

1. Advertease - Real-Time Advertisement Detection

Client Need: A media analytics company needed to detect advertisements in live broadcast streams to verify ad placement and measure screen time.

Technical Approach:

• Built a YOLO-based object detection system trained to recognize TV commercial patterns
• Implemented confidence-scored event emission (probability scores for each detection)
• Optimized end-to-end video processing latency to sub-700ms (critical for real-time use)

Key Challenges:

• Latency Requirements: 700ms end-to-end meant every component needed optimization—video decoding, model inference, post-processing
• False Positive Control: Ads and regular content can look similar. Implemented temporal consistency checks (an "ad" should last 15-30 seconds, not flicker on/off)
• Model Size vs. Accuracy Trade-off: Smaller YOLO models run faster but sacrifice accuracy. Settled on YOLOv5-medium after extensive benchmarking

Outcome: Deployed system processed live streams with 95% detection accuracy and sustained sub-700ms latency. Client used it for ad placement verification across multiple TV channels.

2. Document Processing Pipeline

Client Need: Government contractor processing thousands of scanned forms daily needed automated detection of predefined visual markers (checkboxes, signatures, stamps).

Technical Approach:

• YOLO-based detection pipeline for visual markers in scanned documents
• Data pipeline to process and annotate training data
• Post-processing to extract structured data from detection results

The Data Challenge:

The client provided ~1,000 annotated samples. Not nearly enough for robust YOLO training. I:

• Built data augmentation pipeline (rotation, scaling, noise injection, perspective transforms)
• Expanded dataset from ~1K to ~25K annotations
• Implemented active learning loop—model predictions on unlabeled data, human review of uncertain cases, retrain

Outcome: Achieved ~95% precision in deployment tests. System processed 10,000+ documents per day, reducing manual review time by 80%.

3. OneSKU - Vendor Catalog Alignment

(See dedicated OneSKU project page for full technical details)

Brief Summary: Built hybrid BM25 + embedding retrieval system for aligning heterogeneous vendor catalogs. Achieved sub-15s query latency across multi-million SKU inventories.

4. Sustify - AI-Driven Sustainability Marketplace

(See dedicated Sustify project page for full details)

Brief Summary: Founded an AI-driven sustainability marketplace using transformer-based embeddings for automated vendor matching. Secured early pilot interest before pivoting.

Cross-Project Learnings

Production AI is Different from Research AI

Academic ML focuses on pushing state-of-the-art metrics. Production ML focuses on:

• Reliability: Can this run 24/7 without crashing?
• Latency: Does it respond fast enough for real users?
• Maintainability: Can someone else debug this when I'm gone?
• Cost: Can the client afford to run this at scale?

This mindset shift—from "what's the best model?" to "what's the right model for this production context?"—was transformative.

Data Quality > Model Sophistication

On multiple projects, I learned that:

• Cleaning and augmenting data improved results more than trying fancier models
• Domain-specific data engineering (e.g., separating categorical vs. numerical features in OneSKU) mattered more than hyperparameter tuning
• Active learning loops (model → predictions → human review → retrain) beat passive dataset collection

Latency Budgets Force Trade-offs

Real-time systems teach you about trade-offs fast. For Advertease:

• Could have used larger, more accurate YOLO model → missed latency target
• Could have used smaller model → accuracy too low
• Final solution: medium model + aggressive inference optimizations (TensorRT, batch size tuning, GPU memory management)

Lesson: Understand your constraints (latency, cost, accuracy) upfront and design around them.

Technical Skills Developed

Computer Vision

• Object Detection: YOLO family (v5, v7, v8), Faster R-CNN
• Image Processing: OpenCV, PIL, data augmentation techniques
• Video Processing: FFmpeg, frame extraction, real-time stream handling
• Model Optimization: TensorRT, ONNX, quantization, pruning

Natural Language Processing

• Embeddings: BERT, RoBERTa, Sentence Transformers, OpenAI embeddings
• Semantic Search: Vector databases (FAISS, Qdrant), BM25, hybrid retrieval
• LLMs: Prompt engineering, RAG systems, LlamaIndex

MLOps & Deployment

• Containerization: Docker, Kubernetes for ML workloads
• Model Serving: FastAPI, TorchServe, gRPC
• Monitoring: Prometheus metrics, logging, performance profiling
• CI/CD: Automated testing, model versioning, deployment pipelines

Client Collaboration Skills

Beyond technical skills, PESU Venture Labs taught me critical soft skills:

Requirements Gathering

Clients often don't know what's technically feasible. My job was to:

• Understand the business problem (not just the stated solution)
• Propose technically feasible approaches
• Set realistic expectations (accuracy, latency, cost)

Iterative Delivery

Rather than disappearing for months and delivering a final product, I learned to:

• Deliver MVPs quickly (2-3 weeks for initial prototype)
• Gather feedback early and often
• Iterate based on real user testing

Technical Communication

Clients don't care about model architectures. They care about:

• "Will this solve my problem?"
• "How accurate is it?"
• "What does it cost to run?"
• "When can I deploy it?"

I learned to translate technical details into business outcomes.

Impact & Outcomes

What Made PESU Venture Labs Unique

Unlike traditional university research labs or internships, PESU Venture Labs offered:

• Real clients with real budgets: Accountability to paying customers taught me discipline
• Production deployments: Code actually ran in production, handling real user traffic
• Cross-domain exposure: Worked on computer vision, NLP, and search problems (not just one narrow area)
• Autonomy: Trusted to make technical decisions, not just execute predefined tasks

Transition to Full-Time Work

My time at PESU Venture Labs directly prepared me for my current role at Baxter International. Skills I use daily:

• Translating business requirements into technical solutions
• Building production-ready systems (not just research prototypes)
• Collaborating with cross-functional teams (product, security, DevOps)
• Making pragmatic trade-offs (perfect vs. good enough vs. shipping)