AI systems · developer tools · cloud workflows
I’m Akshay, an engineer in Boston building agentic AI, developer tooling, and workflow automation that stays understandable in production.
$ status --current
building reliable interfaces around model-backed workflows
$ ls selected/
$
# selected work
Public repositories and public-safe case studies, focused on what each system does, where the hard boundary lives, and what the implementation proves.
Vim-style shortcut sequences for Raycast with nested groups, quick capture, browser-aware routing, and portable configurations.
Raycast API · TypeScript · React · Node.js
A conversational booking system connecting planning, live travel search, voice, payments, and observability in one product slice.
FastAPI · Next.js · LangGraph · Stripe
A cloud-native Python platform for extraction, semantic search, summarization, async processing, and API-driven document workflows.
Python · FastAPI · Celery · Vector search
Experiments for isolated agent execution, persistent runtime state, and reusable remote development environments.
Modal · DevPod · Docker · Python
# in production
A full-stack TypeScript platform with payments, background jobs, and an on-site device agent. I own the infrastructure, the delivery pipeline, and the production operations.
Edge runtime
API and SPA served from one Cloudflare Worker: an Express app on the Node compat layer, static assets, per-environment custom domains, and Smart Placement.
Data layer
Serverless Postgres reached through a connection pooler, Prisma on its query-compiler engine, and forward-only migrations checked for drift against a shadow database.
Delivery pipeline
Branch-per-environment promotion, fast-forward only; production deploys only commits already proven on staging. Preflight secret checks, migrate, deploy, health smoke test.
Payments & async work
Idempotent checkout saga, HMAC-verified webhooks with event dedup, acknowledge-then-process, reconciliation sweeps, dead-letter alerts, and a requeue runbook.
Security
Typed env schema resolved from a password vault, scoped tokens per environment, RBAC sessions, CSRF and rate limits, Zero Trust on internal surfaces, and a deploy guard that keeps staging off production data.
Operations
Cut over from a PaaS host to Workers, traced and fixed a ~1.1 MB-per-request ORM memory leak on the edge, and shipped a Windows device agent with remote status and at-most-once commands.
# skills
Python services, data pipelines, and retrieval and agent systems on AWS, measured with evaluation suites and production observability.
Languages & backend
Data & search
AI applications
Cloud & infrastructure
Reliability
Governance
# workflow
How I set up day-to-day engineering: several coding agents in parallel, each fenced off from secrets and production, on environments I can rebuild from code.
Parallel agent harnesses
Claude Code, Codex CLI, and Kimi Code side by side, one config directory per account, in persistent herdr panes inside Ghostty.
Isolation first
OrbStack Linux VMs and Docker Sandboxes microVMs as the boundary per account; untrusted harnesses get repo-only mounts and an egress allowlist.
MCP without leaked secrets
A small launcher every harness calls, pulling credentials from Keychain or a vault per project; secret-free servers via a gateway; a shared fleet over a private network.
Disposable cloud dev servers
Per-account servers defined in code: baked images, persistent home volumes, SSM sessions with no inbound ports, short-lived credentials, and state backups before teardown.
Reproducible projects
Pinned runtimes and package managers, committed lockfiles, pnpm workspaces for monorepos, and typed env schemas instead of .env files.
Knowledge as code
Decisions, evaluations, and runbooks kept as a versioned knowledge bundle with sources and staleness dates, plus HTML explainers and agent briefs.
# work map
The common thread is making powerful systems easier to operate, inspect, and trust.
01 / AI systems
Agent graphs, voice and chat surfaces, semantic search, and practical observability.
See case studies02 / Developer tools
Raycast extensions, browser automation, and explicit shortcuts for repeated work.
View GitHub03 / Cloud environments
Container-first templates and remote sandboxes for agents, builds, and experiments.
Read notes# field notes
Technical write-ups I’m preparing from the systems above.
Hierarchy, capture friction, migration, and browser-aware desktop automation.
plannedWhat to measure across model calls, provider tools, WebSockets, STT/TTS, and payments.
plannedPredeclared functions versus runtime-created containers for isolated agent work.
planned# contact