Case Studies
Documented results across AI systems and cloud infrastructure
A custom speech-recognition engine in production, $160K+ in annual AWS savings, and SOC2 compliance. Client names are withheld, and every number comes from a real engagement.
Custom ASR Engine for a Production AI Platform
Aipa (aipa.io), a commercial AI assistant I build and operate • AI SaaS • In production
Challenge
Real-time voice features needed speech recognition with domain-specific vocabulary accuracy, strict latency budgets, and predictable GPU costs. Off-the-shelf transcription APIs could not meet all three requirements at once, so I built the engine.
Solution
- Built a custom ASR engine on NVIDIA Parakeet (hybrid TDT+CTC), served from a NeMo container
- Implemented decoder-level context biasing: a CTC word spotter grafts domain vocabulary into decoding, with per-word timings and confidence reporting
- Migrated the encoder to bf16 behind strict evaluation gates: word-error-rate, biasing, and false-rejection parity, plus a live A/B on banked audio
- Added TitaNet speaker embeddings (~12ms) for speaker-aware features
- Built a restart-proof audio capture bank for continuous training data collection
Results
- 1.6x throughput on the same GPU (RTFx 257 to 404)
- p50 round-trip latency near 106ms through the serving gateway
- Roughly 2GB of GPU memory freed per instance
- Accuracy parity verified before rollout via live A/B on banked audio
Key Metrics
Technologies Used
Infrastructure Optimization & Cost Reduction
Mid-size Educational Technology Platform • Educational Technology • 6 months
Challenge
Over-provisioned AWS infrastructure with high operational costs and inefficient resource utilization.
Solution
- Implemented autoscaling policies for predictable user patterns
- Refactored RDS configuration to Aurora with reserved instances
- Created automation pipelines for on-demand developer resources
- Executed migration from RightScale to AWS
Results
- Over $100,000 in annual AWS cost savings
- $30,000+ immediate savings from RightScale licensing elimination
- Cut key page response times from roughly 2 seconds to 100ms
- Reduced developer maintenance overhead
- Improved resource utilization efficiency
Key Metrics
Technologies Used
VPC Security & Compliance Configuration
Grant Management Platform • SaaS Platform • 3 months
Challenge
Default AWS VPC configuration inadequate for security compliance requirements.
Solution
- Designed compliant, modern cloud network architecture
- Implemented advanced VPC and subnet configurations
- Created segregated public and private subnet environments
- Established comprehensive security group configurations
- Implemented WireGuard VPN and site-to-site VPN tunneling
Results
- Achieved SOC2 compliance requirements
- Enhanced network security posture
- Enabled secure office-to-cloud connectivity
Key Metrics
Technologies Used
Full-Stack Infrastructure & Frontend Modernization
Consumer Media Platform (embedded principal engineer role) • Consumer SaaS • Multi-year
Challenge
Unstable legacy hosting led to frequent crashes, manual deployments, and inefficient resource use. The front-end codebase had diverged across multiple applications, slowing development and complicating maintenance. Delivered from inside the team as an embedded principal engineer.
Solution
- Replatformed legacy hosting onto containerized, auto-scaling infrastructure (AWS ECS on Fargate)
- Built branch-based CI/CD pipelines for reliable, automated releases
- Secured infrastructure with VPN-only access, private subnets, and centralized controls
- Unified a diverged legacy React codebase onto a single modern architecture
- Managed third-party data-protection assessments and internal audits
- Deployed CDN edge routing for improved global performance
Results
- $60,000+ annual AWS savings
- Stable, scalable containerized infrastructure
- Consistent, modernized front-end stack
- Stronger security posture with segmented access
- Third-party data-protection audits passed
Key Metrics
Technologies Used
Company names are withheld. The Aipa engine is from my commercial product. The other results were delivered in full-time and embedded engineering roles; I now offer the same work as a consultant.