# AI Product Builder Stack

Author: Kevin Astuhuaman
Canonical: https://kevinastuhuaman.github.io/ai-product-builder-stack/
Updated: 2026-07-12

This is a public-safe, evidence-backed map of tools Kevin has used to build and operate AI products. It distinguishes shipped work, active systems, prototypes, and historical experience. It does not disclose credentials, sensitive infrastructure topology, employer data, or proprietary source code.

## Models and agent systems
- Anthropic Claude [active]: Claude SDK and Agent SDK integrations with routed fallbacks, evaluation, and spend controls.
- OpenAI and Codex [active]: Used for model calls, Codex-assisted product building, and the human-reviewed browser-agent harness.
- Amazon Bedrock [active]: Production fallback tier with routing, cost, and failure behavior documented in the backend.
- Azure OpenAI [active]: Trackly uses short-lived WebRTC sessions for its Azure OpenAI Realtime onboarding agent.
- Google Vertex AI [historical]: Provider integration remains documented as historical and disabled, not presented as current production.
- Perplexity [active]: Integrated for research tasks where current web evidence and source attribution matter.
- Model Context Protocol [shipped]: Trackly ships both a public CLI MCP server and a hosted OAuth MCP endpoint.
- Claude Agent SDK [prototype]: Used with explicit tool contracts, limits, and fallback behavior rather than an unbounded chat loop.

## Retrieval and data
- PostgreSQL [active]: The current production database runs on Azure Database for PostgreSQL behind private networking.
- Azure AI Search [active]: Production search combines semantic retrieval with SQL fallbacks and query telemetry.
- LlamaIndex [prototype]: Evaluated for turning uploaded PDF resumes into structured product context.
- Exa [prototype]: Evaluated in hiring-manager and company research flows where semantic web discovery is useful.
- Tavily [prototype]: Configured as a research provider for public hiring signals and source-backed enrichment.
- Parallel [prototype]: Used in research experiments; not represented as a production dependency.

## Browser and automation
- Codex computer use [prototype]: The agent fills and verifies fields, then pauses before submission for fresh human approval.
- Custom agent harness [prototype]: Encodes ATS mechanics, committed form state, recovery, approval invalidation, and confirmation handling.
- Puppeteer [active]: Used in the backend for browser-based ATS coverage and controlled automation.
- Playwright [active]: Runs end-to-end tests and visual checks for Trackly Web and public product studies.
- Chrome DevTools Protocol [shipped]: Used in a public X article publisher and browser-control workflows.
- Human approval gates [shipped]: Applied to browser submission, generated follow-ups, and evidence-aware investigation patterns.

## Infrastructure and delivery
- Microsoft Azure [active]: The live Trackly backend and data plane run on Azure after a documented cloud cutover.
- Azure App Service [active]: The current Azure deployment separates user-facing, API, and background workloads with guarded CI paths.
- Azure Key Vault [active]: Included in the reviewed production infrastructure inventory without publishing values or access topology.
- Terraform [active]: Azure production and AWS rollback resources have reviewed, non-mutating Terraform plans.
- AWS Lightsail [historical]: Modeled in the cloud migration inventory as historical infrastructure, not current production.
- Amazon RDS [historical]: Documented in the audited AWS-to-Azure migration rather than presented as live production.
- Amazon SES [active]: Used behind email generation and follow-up operations with deliverability controls.
- Cloudflare [active]: Supports Trackly domains and the recruiter portfolio delivery path.
- Docker [active]: Used for backend delivery and multi-service public architecture demonstrations.
- GitHub Actions [active]: Runs multi-repo CI and the guarded deployment paths for web, backend, native, and public labs.
- Vercel [historical]: Used for earlier product surfaces before the current Azure and GitHub Pages paths.
- Railway [historical]: Retained as historical builder experience, not current infrastructure.

## Observability and analytics
- Langfuse [active]: Instrumented across model routes with graceful failure behavior and public architecture notes.
- PostHog [active]: Tracks funnels and product decisions across Trackly surfaces with privacy-aware event design.
- Umami [shipped]: A public Go MCP server turns Umami analytics into structured assistant tools.
- Gatus [active]: Deployed as part of the current public-surface monitoring path.
- Healthchecks [active]: Used to detect silent failures in recurring operational workflows.
- Firebase [active]: Integrated into Trackly iOS and macOS release builds.
- Datadog [prototype]: Used during the PayPal Checkout internship prototype; no employer data or architecture is published.
- Amazon CloudWatch [historical]: Documented as experience from the AWS operating period, now secondary to Azure monitoring.
- Application Insights [active]: Monitored as part of the production public-surface and infrastructure drift checks.
- Azure Log Analytics [active]: Included in the current observability inventory while keys and private workspace details remain excluded.

## Voice and communication
- Azure OpenAI Realtime [shipped]: The native apps use ephemeral tokens, direct audio transport, tool calls, and session observability.
- WebRTC [shipped]: Implemented in Swift with reconnect handling, audio routing, and sanitized transport metrics.
- Whisper [shipped]: Trackly includes a Whisper-based transcription path in its iOS product.
- Fireflies [active]: Webhook integrations turn completed conversations into structured drafting tasks.
- Hume AI [prototype]: Used to test whether emotional signals can improve follow-up quality and recall.
- Twilio [active]: Connects background processing to time-sensitive human review.
- Vapi [historical]: Kept as historical product exploration rather than presented as the current implementation.

## Product engineering and quality
- Next.js and React [shipped]: Trackly Web ships responsive product surfaces, agent chat, onboarding, and job management.
- SwiftUI [shipped]: Trackly ships separate native applications with shared product contracts and platform-specific UX.
- Astro [shipped]: Powers the investigation workbench, recruiter portfolio, and this evidence map.
- TypeScript [active]: Used across production services and public-safe prototypes with automated type checks.
- Go [shipped]: The public server includes tests, schemas, container packaging, and release automation.
- CodeRabbit [active]: Used with CI, Codex, security review, and explicit resolution rather than treated as a sole approval.
- Conductor [active]: Used to coordinate multiple bounded branches while preserving review and merge gates.
- Figma [prototype]: Used for Berkeley and MoBagel AI product concepts, UX exploration, and roadmap communication.
- Remotion [shipped]: Used to ship a public Trackly CLI product video with repeatable, code-reviewed motion.
- Xcode and TestFlight [shipped]: Used for Trackly iOS and macOS delivery, including guarded builds and TestFlight distribution.
- Replit [historical]: Included as historical builder experience rather than a current delivery dependency.

## GTM and operating tools
- Clay [active]: Used in relationship and outreach workflows with explicit enrichment quality checks.
- HubSpot [active]: Used for contact management and source attribution workflows.
- Ghost [shipped]: Used in a voice-to-newsletter workflow and Kevin's publishing system.
- Gmail and Calendar [active]: Integrated through OAuth with approval-oriented relationship workflows.
- Notion [active]: Used in integration and workflow experiments where editable human context matters.

Structured data: https://kevinastuhuaman.github.io/ai-product-builder-stack/stack.json
Portfolio: https://portfolio.kevinastuhuaman.com
GitHub: https://github.com/kevinastuhuaman