Echios
PythonAWS Bedrock (Claude)AWS TranscribeDeepgramPyTorchscikit-learnOpenSearch ServerlessDynamoDBLambdaS3CloudFormation/SAM
Echios is a desktop application that helps public company executives prepare for and deliver more effective earnings calls through real-time speech analysis, AI-generated coaching, and predictive question modeling. Think Poised or Yoodli, but purpose-built for investor relations, with domain-specific metrics like forward-looking statement detection, analyst question prediction, and answer consistency tracking.
Role: Principal Software Engineer / Core Architect
Journey: Proof-of-Concept → MVP Production Release
Leadership & Architecture
- System Architecture: Designed a layered, modular architecture (~50K LOC) with clean separation across 15+ subsystems: audio capture, streaming transcription, real-time speech analytics, LLM orchestration, vector-based RAG, and a desktop UI, all coordinated through async event-driven patterns with sub-500ms latency targets.
- Technical Decision-Making: Made critical technology bets: Tkinter for desktop-first real-time performance over web frameworks; AWS Bedrock (Claude) over OpenAI for enterprise alignment; dual ASR providers (AWS Transcribe + Deepgram) for resilience; in-memory vector stores for fast iteration with OpenSearch Serverless for production scale.
- PoC → MVP Execution: Drove the product from initial prototype to production release, standing up CI/CD pipelines, platform-specific builds (py2app for macOS, PyInstaller for Windows), AWS infrastructure-as-code (CloudFormation/SAM), and a comprehensive test suite (~19K LOC) spanning unit, integration, and platform-specific tests.
- Cloud Infrastructure: Architected the serverless backend on AWS: Lambda + API Gateway for auth, DynamoDB for user management, STS for ephemeral credential exchange, S3 for audio storage, and OpenSearch Serverless for production vector search.
Key Technical Achievements
- Real-Time Audio Pipeline: Multi-source audio capture (microphone + system audio via ScreenCaptureKit on macOS, WASAPI on Windows), stereo mixing, and streaming to ASR, all at 16kHz with 100ms chunk buffering.
- Live Speech Analytics Engine: Built a suite of real-time analysis modules: disfluency detection (fillers, hedges, repairs), clarity scoring, pace tracking, rambling detection, and financial terminology tracking, processing transcript tokens as they stream in.
- LLM-Powered Coaching: Integrated Claude via AWS Bedrock for predictive analyst question generation from earnings scripts, real-time answer quality evaluation, context-aware talking point retrieval via RAG, and an "LLM Judge" quality gate on generated analysis.
- Knowledge Base & RAG: Document ingestion pipeline (PDF, DOCX, PPTX, XLSX) with chunking, Titan embedding generation, and semantic retrieval powering contextual coaching across practice, live, and manual modes.
- Cross-Platform Shipping: Delivered standalone desktop builds for both macOS and Windows with embedded Python runtimes, platform-specific audio backends, and automated CI/CD build pipelines.