Architecture of the systems I build and operate.

Architecture of Harsh's personal pi-coding-agent delegation setup: one frontier executor that owns every write, fanned out across specialist lanes — a read-only advisor for judgment consults, cheap parallel scouts for recon, and sandboxed coders for mechanical implementation — with two enforcement extensions (advisor-autoconsult, verified-writes) that put the guardrails in the harness rather than the prompt.

Architecture of the agentic build-failure triage system on Memex: a read-only Jenkins connector (the tool layer) plus a markdown runbook (the skill layer), bound by a signal→action table, a bounded-log tokenomics access pattern, and a structured root-cause report contract.

Architecture of the agent-facing code-search tool on Memex: a Zoekt trigram index over EFS-mirrored git repos, glob→regex translation at the tool boundary, a return-matches-not-files formatter, nightly indexing, and an arm64 zoekt-webserver fronted by an nginx sidecar.

Architecture of Open Brain — Harsh's persistent, cross-session semantic memory layer exposed to any AI agent over MCP (HTTP/SSE): a Supabase edge-function logic plane (an MCP server plus a Slack-integrated ingest function), OpenRouter AI services for embeddings and metadata extraction, and a Postgres+pgvector data plane holding a thought stream and a taste-preference registry.

Architecture of a personal productivity system where Claude Skills are the procedures and markdown files are the database: a thin router (CLAUDE.md) dispatching to three sub-agents — Planner, Scheduler, Reviewer — over a gitignored data layer of weekly sprints, themed project files, an inbox, and a backlog, with a shared conventions reference and a model-tier policy that runs cheap models for tactical work and frontier models for judgment.

Architecture of the release platform Harsh owned for seven years: stage-driven Jenkinsfiles that drive an issue tracker's release lifecycle, a shared-library + job-DSL two-layer design, and the trunk-based daily-release cadence it enabled — with the friction that shaped it kept as seasoning.

Architecture of an agentic archival pipeline for recurring multi-source research: dated, immutable, never-overwritten capture runs that preserve raw originals verbatim and refuse to analyze unless explicitly asked; browser-auth session reuse with an ask-before-continuing rule; a per-run folder layout of raw connector JSON, content extracts, originals, and download manifests; config-driven source scope; and deliberate restraint — a loose archive with no git and no app.

Architecture of a local Next.js dashboard that turns per-assistant AI usage into one scrubbed daily-burn row: exact local telemetry where logs expose token counters (Codex, Claude Code), conservative request-count estimates where they don't (GitHub Copilot), Pacific-day bucketing, and a driver annotation that names what drove the burn — with no content ever leaving the machine.