Work

Inspection AI

Applivity · RealWear First Agentic Field Inspection Platform

inspectionai.applivity.com ↗

• Own the architecture for Inspection AI, an agentic, voice-driven inspection platform for RealWear devices, combining a .NET Core backend with a React/TypeScript frontend deployed to Azure App Service via GitHub Actions CI/CD.
• Indexed inspection manuals, prior reports, and reference documents in Azure AI Search so field technicians can query relevant documentation hands-free while an inspection is in progress.
• Combined BM25 keyword ranking and fuzzy search with vector similarity to improve retrieval precision on domain-specific terminology, part numbers, and misheard voice transcriptions.
• Applied prompt engineering, including few-shot examples and structured output constraints, to turn retrieved document passages into concise, voice-ready answers during inspections.
• Set the authentication and access model: secure RESTful APIs with JWT-based auth and RBAC, plus end-to-end OAuth2/OIDC flows with token refresh across a multi-tenant user base.
• Led research and prototyping for a fully offline, on-device speech pipeline for RealWear hardware: a new realwear-app module (Kotlin, Jetpack Compose, minSdk 26, targetSdk 35, arm64-v8a only) running sherpa-onnx models locally, with the legacy app module untouched.
• Designed the offline model catalog (Whisper tiny.en for STT at roughly 117MB, and tiered Piper TTS voices at roughly 20–33MB) with tier-aware model selection and a pluggable response strategy so the pipeline can run on echo/rule-based logic today and swap in an LLM later without a redesign.
• Evaluated candidate STT and TTS models against latency and quality metrics, including tokens per second, time to first token, and word error rate, and built retry and failure handling so the on-device pipeline degrades gracefully rather than hanging.

Incident AI

Applivity · Voice Driven Field Incident Reporting

incidentai.applivity.com ↗

• Led delivery of a production voice-driven incident reporting system for field engineers in remote oil and gas environments (ADNOC Drilling), integrating the Azure Speech SDK for real-time speech-to-text and text-to-speech across 9+ languages.
• Cut incident logging time from several hours to under 2 minutes by replacing manual forms with voice-driven data capture routed through structured backend APIs.

Video Analysis AI

Applivity · Training Content Search Pipeline

• Built a video analysis pipeline for internal training content: incoming footage is chunked into short indexed segments and labeled with structured metadata and precise timestamps for fast retrieval.
• Indexed chunked video segments in Azure AI Search and Pinecone to support both keyword and semantic search, letting trainers jump directly to the relevant moment across a large video library instead of scrubbing full recordings.
• Tuned the chunking and labeling strategy specifically for training use cases, balancing segment length against search precision so results stay tightly scoped to what a trainer is looking for.

PDF & Document Intelligence Pipeline

Applivity · Multi Agent AI Architecture

• Designed a dual-agent AI architecture: an Extraction Agent that parses unstructured documents (forms, invoices, field reports) via prompt-engineered LLM calls, and a Verification Agent that cross-validates extracted fields against business rules and flags anomalies for human review.
• Built the extraction pipeline on Azure Document Intelligence (Form Recognizer) for high-accuracy parsing across varying document layouts, feeding downstream LLM reasoning and RAG-based lookups.
• Applied prompt engineering techniques, few-shot examples, chain-of-thought reasoning, and structured output constraints, to raise extraction accuracy and materially reduce hallucination rates across document types.

Rapticore

Cloud Security & Data Governance SaaS Platform

rapticore.com ↗

• Designed multi-cloud data acquisition pipelines (AWS, GCP, Azure SDKs) feeding an automated, always-current data governance catalog used to power downstream AI-driven risk scoring.
• Engineered a Neo4j graph database with FastAPI REST APIs to model cloud resource relationships, delivering a 95% performance improvement over relational approaches for lineage and dependency queries.
• Built a compliance rule engine (CIS AWS Benchmark v1.4 plus custom rulesets, 50+ automated checks) that flags misconfigured IAM roles, exposed data stores, and access key violations in real time.
• Designed an event-driven ingestion architecture (AWS EventBridge and SQS), cutting detection latency from hours to under 10 seconds across distributed account fleets.
• Directed deployment of containerized microservices (Docker, AWS ECS) and infrastructure provisioning via Terraform and AWS CDK, reducing enterprise onboarding time by over 60%.

Engineering Memory OS

CockroachDB × AWS Hackathon · Aug 2026

• Leading a 2–3 person team building an AI-powered institutional memory platform that integrates GitHub, GitLab, Slack, and Jira to automatically capture engineering decisions and answer natural-language questions such as why a decision was made, using persistent AI memory.
• Owning the architecture for the retrieval/memory layer on CockroachDB alongside a RAG-based query engine, with MCP-based connectors for pulling structured context from engineering tools in real time.
• Scope spans LLM/agent design, vector retrieval, AWS deployment, and full-stack delivery, end-to-end from architecture through demo.

Client Facing Software Engineer

Remotebase · Jul 2023 – Jan 2024

ember.new ↗

• Worked directly with product managers and clients to translate business requirements into working software on the Snapdev platform (now Amber.ai), serving as primary technical point of contact through iterative delivery.

Full Stack Engineer

Thathal I&TS · Sep 2023 – Aug 2024

firststopdental.com ↗

• Built Google Ads API connectors for automated campaign data ingestion and real-time analytics, plus React/TypeScript dashboards for performance monitoring.
• Co-developed a microservices-based enterprise SaaS platform with containerized deployments, cutting deployment time by 40%.