Industry

CVision

• Co-founded CVision and led AI-first automation for clients across the US and EU; built the founding engineering team and shipped MVPs and production systems generating $10K+ in first-month revenue.
• Led technical design and system architecture for 10+ paid client engagements and pilots across SME operations, logistics, supply chain, and private-equity-backed businesses.
• Designed and deployed applied AI for inbound/outbound automation, operational intelligence, and decision support—systems that directly enabled $2M+ in client-side revenue.
• Architected knowledge-graph-based memory systems linking real-time and historical data for causal analysis in logistics and supply chain operations.
• Scaled the company to $80K ARR in under 1 year, grew the engineering team to 8 people, and operationalized AI-assisted development workflows across the organization.
• Led technical vision, roadmap, and architecture for constrained, industrial, and non-cloud-first environments; drove a strategic pivot from generic AI automation to deep-tech and efficient computing.

Upwork (Freelance)

• Designed and deployed AI automation pipelines that reduced manual operational effort by up to 70% for clients in operations and logistics.
• Scoped, built, and delivered 3+ production-ready AI systems end-to-end—from discovery and architecture through deployment, documentation, and stakeholder handover.
• Advised early-stage founders on AI adoption strategy, technical feasibility, and build-vs-buy decisions for live business environments.

Cowlar Design Studio (YC W17)

• Deployed retail pricing automation on edge devices using Qualcomm Hexagon NPU for millisecond-level inference, outperforming GPU speeds.
• Built a platform for automating vision applications with active learning workflows, streamlining ML data collection, annotation, and model deployment.
• Automated factory operations for ViaPhoton, an Amazon-partnered fiber optics supplier, boosting production capacity by 40x.
• Built an ML/CV system achieving 96% precision at 5μm alignment, scaling monthly output from 240 to 9,600 cables.
• Enhanced the Fiber Organizer Robot using CV and ML to sort 250μm fibers, increasing output by 500% (1 to 6 AWS cables/hour).
• Redesigned an action-recognition model for smart carts based on Mobile Video Networks, reaching 95% real-time edge inference accuracy.

Anthropic

• Designed and operated an RLHF evaluation pipeline for Claude 3.5 Sonnet and Haiku, targeting adversarial and edge-case coding tasks to stress-test model reasoning.
• Produced structured preference rankings and fine-grained feedback used to improve post-training signal quality for code generation.
• Led weekly calibration and failure-analysis sessions with the evaluation team, contributing to a 15% lift in coding benchmark performance.

Optical Networks and Technologies Lab

• Implemented LSTM, GRU, and CNN models for temporal signal-processing tasks in optical-network fault prediction and performance monitoring.
• Prototyped attention and transformer-based architectures for anomaly detection, improving interpretability of failure modes in optical signals.
• Applied GAN-based synthetic data augmentation to mitigate data scarcity, improving model robustness under noisy real-world conditions.
• Built reproducible experimentation workflows with version-controlled experiments and systematic hyperparameter search.