Namaste Ji by Ayushman Dash

Docs / roadmap.md · mirrored from the repo

Roadmap & Personal Todo

Personal north-star list. Update as phases complete.

Phase 1 — Events Calendar Agent + Agent Tooling Foundation

Goal: ship the first agent end-to-end with full visibility and control. Establish the recipe every future agent will follow.

1a. Events Calendar Agent (core)

  • Content catalog data model — D1 schema + R2 layout + Vectorize index (shared spine, do first)
  • Control DO interface — modes (autonomous / manual / paused), decisions log, audit trail
  • Implement the Agent Kernel in agents/events-calendar/ (STM + LTM via MemoryPort, calibrated-stop loop, subagent-as-tool firewall)
  • Wire all prompts through Langfuse (name@label, versioned) — no inline prompt strings
  • Subagent roster: web-research · catalog-research · code-exec · memory · strategy-read
  • Cron Trigger + manual /run endpoint
  • Full Cloudflare AI Gateway routing for every model call (local + staging + prod)

1b. Back-Office tooling (must ship alongside the agent)

  • Run control UI — trigger a run, pause/resume, set mode (autonomous / manual / scheduled)
  • Run history list — every run with status, timestamps, token cost, model used
  • Run detail view — step-by-step trace: every tool call, every LLM call, input + output at each step
  • Langfuse integration — prompt traces linked from run detail; score/label runs from the BO
  • AI Gateway dashboard link — deep-link to the matching gateway log from each run
  • Control DO panel — view current mode, flip autonomy level, see pending decisions needing human review
  • Config / Settings page — edit KV-backed config live (model_id, provider, schedule, thresholds)
  • Rerun / replay — trigger a rerun of any past run, optionally with edited inputs
  • Durable Objects run log — store run metadata in DO or D1 so it survives Worker restarts

1c. Notification service

  • Design the notification contract (event types: run-started, run-complete, run-failed, human-review-needed, cost-alert)
  • Email notifications (Resend or Cloudflare Email Workers) — at minimum for errors + review gates
  • In-app push (BO console — browser push or polling banner) for real-time run status
  • Notification preferences in BO Settings (per-event, per-channel on/off)

1d. Agent recipe / template

  • Write docs/agents/AGENT-RECIPE.md — a checklist a new agent must satisfy before it’s considered “done”
  • Extract reusable Agent Kernel code into packages/agent-kernel/ so future agents import it
  • Document the BO wiring contract: what every agent must expose (run endpoint, status, config keys, DO log schema)

Phase 2 — Creative Head + Strategy Agents

Build on the recipe. The events-calendar agent’s tooling should make these faster to ship.

  • Creative Head agent (agents/creative-head/) — thin orchestrator, single versioned Content Strategy artifact
  • Subagent roster: catalog-research · code-exec · web-research · memory · strategy-read · content-sampler
  • Cold-start bootstrap flow (empty catalog path)
  • Content Strategy artifact versioning in D1 + R2 (strategy@vN)
  • Sampler probes → STM feedback loop
  • Human-in-the-loop gate in BO for strategy approval before production runs
  • All other creative / content agents (TBD as Creative Head design matures)

Phase 3 — Eval Components

Don’t build evals until the agents are producing real output — eval design should be grounded in actual failure modes.

  • Define eval dimensions per agent (correctness, novelty, cultural fit, vernacular quality, …)
  • Judge agent (LLM-as-judge) with Langfuse annotation queues
  • Golden dataset (first batch from Phase 4 becomes the seed)
  • Automated eval pipeline on each run (score → Langfuse → BO dashboard)
  • Human annotation workflow in BO (batch labelling UI)

Phase 4 — First Data Generation Run

Use the full agent stack to generate the first real content batch.

  • Run the Creative Head end-to-end on staging
  • Human review in BO — approve / reject / edit content
  • Ingest approved content into the content catalog (D1 + R2 + Vectorize)
  • Validate catalog data model against real content shape
  • Capture eval scores; identify top failure modes; feed back into agent prompts

Phase 5 — Frontend

Decide stack after agents are proven — don’t over-engineer before then.

  • Decide: React Native (cross-platform mobile) vs Flutter vs PWA vs web-first
    • Android-first given the audience; “build once, distribute across platforms” is the goal
  • Consumer auth — Better Auth on Workers (phone OTP + Google)
  • Content feed UI — server-side personalisation, edge delivery
  • Share intent signal (light, safe) wired to distribution plane
  • Notification delivery to end users (push / WhatsApp / SMS TBD)

Phase 6 — UX Iteration + Early Users

  • Onboard parents + in-laws (first 4 users)
  • Collect qualitative feedback — what do they actually share? What do they ignore?
  • Iterate on UX based on real usage patterns
  • Define “ready to onboard strangers” criteria (content quality bar, auth, basic moderation)
  • Soft launch

Cross-cutting (always on)

  • Keep docs/ARCHITECTURE.md + AGENTS.md status snapshot current after every phase
  • Every agent PR must have human review before merge (CODEOWNERS → @dashayushman)
  • Secrets never in repo — rotate any token accidentally committed immediately
  • Alchemy IaC (infra/) stays the single source of truth for CF substrate