I design LLM, RAG, agentic, and voice solutions — then run them on AWS with Terraform. Shipped end to end for clients in e-commerce, healthcare, and retail, from the first prototype to a live Marketplace listing.
Designed and shipped a real-time STT–LLM–TTS pipeline with multi-step agentic workflows on Strands Agent and the Agno framework, published as a package on PyPI. Cut manual support effort by roughly 40% and response times by 35%.
Owned two voice agent systems for a US live-commerce brand — wired into the e-commerce backend for order management, complaint resolution, Jira ticketing, and live agent escalation, then deployed across their telephony stack.
Built an OCR-to-RAG pipeline: AWS Textract for layout-aware extraction, OpenAI and sentence-transformer embeddings indexed in ChromaDB, retrieval served back through an LLM so enterprise sites can hold a conversation with their own unstructured documents.
A clinical platform on AWS HealthScribe that transcribes live doctor–patient conversations and extracts structured notes — less time spent on documentation, cleaner EHR data.
Contributed to the company's main AI demos website and proposed new AI product concepts mapped to client verticals — turning delivery patterns into things we could sell again.
Published production AI solutions to AWS Marketplace — infrastructure built for enterprise-grade security, auto-scaling, and high availability, not a demo that falls over at the second concurrent call.
Every environment provisioned through Terraform and AWS CloudFormation — dev and production stand up from the same version-controlled definitions, so deploys are repeatable, auditable, and reversible. Infrastructure reviews in a pull request, like the rest of the codebase.
Compute, storage, data, and inference sized and wired together: EC2, S3, RDS, Lambda, Bedrock, plus Amazon Connect contact flows and IVR logic carrying real-time inbound and outbound telephony at scale through Twilio and Telnyx.
AWS Certified Solutions Architect – Associate (Jan 2025) and HashiCorp Certified: Terraform Associate (Feb 2026) — the practice came first, the certifications confirmed it.
A full speech-to-speech engine using local STT, in-house text processing, and TTS for low latency and contained data. Agentic workflows handle order lookup, complaint handling, Jira tickets, appointment booking, outbound calls, and human handoff. Deployed on AWS Marketplace.
View on PyPISpecialist agents collaborating over Agno Teams with agent-to-agent delegation, shared memory, and central orchestration. Dynamic prompt chunking spreads large inputs across agents without losing task context — adaptive routing and context compression fixed the sequential-invocation latency wall.
Deepgram TTS against Minimax TTS for real-time voice agents, with Amazon Polly as the quality baseline. Python benchmarking pipelines and dashboards scored latency, streaming behaviour, naturalness, multilingual coverage, and cost — the numbers picked the engine.
Schema-aware prompting on Amazon Bedrock (Claude, Titan) translates plain questions into executable SQL and MongoDB queries — joins, aggregations, filters, nested documents. Results render as tables and charts, so reporting no longer waits on an analyst.