Deep Dive: Okwel Edgar Mark - Optimization, STT Pipeline & Security
How Okwel Edgar Mark optimized performance across the stack, built the speech-to-text pipeline, and implemented the security guardrails protecting taxpayer data.

Okwel Edgar Mark
Deep Dive: Okwel Edgar Mark - Optimization, STT Pipeline & Security
Introduction
Okwel Edgar Mark made the URA Chatbot fast, gave it ears, and kept it safe. His work spans performance optimization, the speech-to-text pipeline, and the security implementation that protects sensitive taxpayer information.
Optimization
Edgar tuned the system for speed and efficiency:
- Caching strategy: semantic and response caching to avoid repeated work
- Latency reduction: profiling hot paths and trimming overhead
- Quantized inference: 4-bit model serving for memory efficiency
- Right-sizing and auto-scaling: matching resources to real demand
The result is sub-2-second average responses that hold up under peak load.
The STT Pipeline
So taxpayers can simply speak their questions, Edgar built the speech-to-text pipeline:
- Automatic Speech Recognition for English and Luganda
- Whisper-based models with adapters for local languages
- Noise handling for real-world phone audio
- Streaming transcription feeding straight into the RAG engine
Security Implementation
Edgar built security into every layer, not as an afterthought:
OWASP LLM Top 10 safeguards
- LLM01: Prompt-injection detection
- LLM02: Insecure output handling
- LLM06: Sensitive-information disclosure prevention
- LLM08/09: Excessive-agency and overreliance mitigation
Platform security
- TLS 1.3 in transit and AES-256 at rest
- PII detection and redaction
- Input validation, rate limiting, and audit logging
Impact
Edgar's work delivered:
- Sub-2-second responses sustained under load
- Voice input in English and Luganda via the STT pipeline
- 100% of simulated attacks blocked during security testing
- A system taxpayers can trust with sensitive information