Akshit Garg 

Most "AI content tools" are a single prompt wrapped in a UI. This one is built like a real editorial pipeline: a planner decides the angle and hook, a generator writes the draft, a fast rule-based scorecard filters weak output before it ever reaches an LLM critic, and an adversarial critic agent reviews it against six quality dimensions before a human ever sees it — with automatic retry and replan loops when a piece doesn't clear the bar.

• Multi-agent orchestration with real failure handling — every generation step has retry/replan logic with capped attempts, and every LLM call degrades gracefully to a stub instead of crashing the pipeline.
• Fast filtering before expensive calls — the virality scorecard is deliberately rule-based, not an LLM call, so weak drafts get killed before burning a critic-model request.
• I own the infrastructure, not just the app — I designed and deployed the full production stack myself: Cloudflare Tunnel (zero exposed ports, no port-forwarding), Redis and Qdrant running natively on a self-hosted Mac mini, Postgres via Supabase, all behind a hardened firewall with every internal service locked to loopback.
• Shipped, not just scaffolded — resolved real production dependency conflicts across pydantic/httpx/supabase, applied 20 sequential database migrations, and cut over a stale deployment to 8,000+ lines of unreleased work with zero functionality lost.

The interesting constraint here was deliberately not training a model. The source material — a year of a nutritionist's real coaching — is far too small for fine-tuning and doesn't need it. So the method is captured as an explicit ruleset plus worked examples and replayed at inference time. The result is auditable: every plan can be traced back to the rule that produced it, which matters a great deal more in health than it does in content.

• Expert method as rules, not a trained model — retrieval plus a grounded prompt reproduces the nutritionist's three-signal loop (weight trend, standing rules, adherence). Verified end to end: on a gaining trend with logged deviations, the generator independently selected the correct reset protocol and cited the specific slips that triggered it.
• Time-aware trend, so stale data can't do harm — only weigh-ins inside a configurable window count. A user returning after a break gets a gentle onboarding rotation instead of an aggressive correction computed off months-old numbers — a failure mode that would have been actively counterproductive.
• Health data isolated at the database layer — per-user row-level security in Postgres, so isolation is enforced by the database rather than trusted to application code. The admin dashboard is the one path that bypasses RLS, behind a service-role client restricted to allowlisted admin routes.
• Owned end to end, again — built, deployed, and self-hosted on the same Mac mini as Authentic: launchd-supervised Next.js on loopback, reverse-proxied through Caddy, exposed only via Cloudflare Tunnel with zero open ports.