Akash Ramesh

Personal voice assistant: a Python orchestrator driving Claude tool calling, woken by a custom trained "Hey Ash" wake word.

• Tool calling loop on the Anthropic Messages API executes tool requests and feeds results back until the model produces a final answer
• Execution and personality kept separate: tools run literally, a versioned persona spec styles the reply, destructive actions always need confirmation
• Custom "Hey Ash" wake word model trained on synthetic speech with adversarial negatives and noise augmentation, exported to ONNX
• Personalisation layer injects active project, standing facts, and recent commands into every request
• Fully local speech stack keeps voice at zero marginal cost, estimated AUD $10 to 20 a month in total

Service	Role
Claude API	Reasoning and tool calling orchestration
openWakeWord	Custom "Hey Ash" wake word detection
Whisper	Local speech to text
Piper	Local text to speech
AWS Lambda	Cloud orchestrator for all devices (planned)

Photo based repair approval for car workshops: the mechanic texts a quote link, the customer approves from their phone browser with no login or app.

• Application developer on a 3 person team, owning the mechanic app and the customer report page
• Customer links secured by a single use, high entropy token hashed before storage, so a database leak cannot produce working links
• Browser uploads straight to a private S3 bucket via short lived presigned URLs, the API never touches file bytes
• Found and fixed an iOS Safari bug where WebP silently fell back to PNG, making compressed photos larger
• Append only events table tracks time to decision and approval rate from the first customer link

Service	Role
AWS RDS	PostgreSQL with transactional, row locked writes
AWS S3	Private photo storage via presigned URLs
GitHub Actions	CI with Playwright on mobile Chrome and Safari
SMS provider	Quote link delivery (planned)

Multi tenant SaaS portal wrapping an enterprise search API, shipped to a paying client.

• Stripe metered billing across three usage dimensions with idempotent usage reporting to prevent double counting
• Keycloak 26 Organisations for multi tenant isolation, active_tenant claim resolved on every backend request
• ASP.NET Core backend proxy pattern keeps BookStack and Zammad PATs server side, never exposed to the browser
• Onboarding reduced from manual provisioning to under 5 minutes self service
• Shipped v1.0.0 to a live paying client

Service	Role
Keycloak 26	Identity and multi tenant org management
Stripe	Checkout, subscriptions, metered billing
BookStack	Knowledge base via backend proxy
Zammad	Support ticketing via backend proxy
Azure App Service	Hosting for containerised services

iOS app for building and reusing AI personas as structured system prompts.

• Temperature computed from tone base value plus a consistency slider: final temp = base + (consistency - 0.5) x 0.4, clamped to 0 to 1
• Six tone presets (Precise 0.1 to Friendly 0.6) map user intent to LLM temperature without exposing raw values
• OpenAI used to generate structured JSON persona suggestions from plain language descriptions, decoded directly into the Persona model
• Two tier persistence: SwiftData for local drafts, Firestore for cloud sync per authenticated user
• WidgetKit extension reads favourite persona from App Group UserDefaults for instant home screen access without a network call

Service	Role
Firebase Auth	Sign in with Apple and Google Sign In
Firestore	Cloud sync for finished personas per user
OpenAI API	Persona suggestion and prompt generation
WidgetKit	Home screen widget for quick persona access
App Groups	Shared UserDefaults between app and widget

Freelance migration from a no code builder to a self hosted WordPress CMS for a healthcare practice.

• Full end to end ownership: DNS, hosting setup, WordPress installation, theme build, plugin development, and client handover
• Custom theme and plugin set built so non technical staff can manage all content without developer involvement
• Migration carried zero additional hosting cost by running WordPress on the existing VentraIP plan
• Elementor page builder layer gives staff layout control within brand constraints set by the custom theme
• Next phase scoped: HotDoc booking integration and Claude API powered FAQ chatbot

Service	Role
VentraIP	Shared hosting, DNS, cPanel management
Elementor	Page builder for staff layout control
HotDoc	Online booking integration (planned)
Claude API	AI powered FAQ chatbot (planned)

Backend microservice for direct income capture at card activation for a global banking client.

• Replaced a 3 to 6 month manual review process with real time income capture at the point of card activation
• Reached 61% customer adoption in the initial release
• Cut external data vendor costs by 20% and lifted offer conversion by 12%
• Extended existing Hadoop and Hive pipelines to feed captured income data to Tableau dashboards without new tooling
• Apigee gateway handled auth and rate limiting, keeping the microservice focused on business logic

Service	Role
Apigee	API gateway, auth, rate limiting
Jenkins	CI/CD build and deployment pipelines
Hadoop/Hive	Batch pipeline to enterprise datalake
AWS ECS	Container hosting post PCF migration
Tableau	Business analyst dashboards via datalake