Full-stack engineer, 7+ years deep — React & React Native up front, event-driven Node/TypeScript underneath. These days I point that stack at agent systems that survive production: queues that hold, sends that can't go out half-formed, human checkpoints where they matter.
This is the actual shape of my job-hunt agent — and of every agent I ship. Deterministic filters before any model call, batch scoring to amortise prompts, and a human gate before anything irreversible. Watch the tokens: nothing crosses a checkpoint on its own.
Every card carries a live miniature of the system it describes — because "trust me, it works" is not an engineering argument. Agents up top; the full-stack products underneath them — storefronts, dashboards, mobile apps — further down.
An agent that runs a senior job hunt end-to-end: sources roles, scores them in batch, drafts outreach — and learns which voice gets replies via a Thompson-sampling bandit. Human approval over Telegram at every send.
The control plane behind my product suite: per-project FIFO queues, job lifecycle with cancel scoping, state-preserving releases with WAL checkpoints — v0.14.8 after a structured 7-bug audit cycle.
Pulls assigned bug tickets, proposes a fix, applies it, opens PRs against upstream and base — with a human checkpoint at every step. The agent proposes; a person stays accountable for what lands.
A reel-rendering monorepo with a headless export worker: Playwright captures an image sequence frame-by-frame, ffmpeg assembles with HDR→SDR tonemapping, BullMQ keeps cancellation honest across tab unloads and restarts.
Jewellery photo studio in the browser: multi-engine background removal (RMBG-1.4 / BEN2), brand overlays on a Konva canvas, plan-gated catalog PDF export. Drag the divider — raw phone shot to studio-grade.
Multi-channel outreach across WhatsApp, Instagram and email with fuzzy identity linking, a conversation-scoring engine and a Hinglish-aware sentiment classifier — because real Indian B2B conversations don't happen in textbook English.
A jewellery retailer's storefront end to end: SSR product pages, live gold-rate ticker, wishlist, plans, cart → checkout → orders — public/protected routes split via HOC + dual API clients, per-route SEO with structured data.
Jewellery management for retailers: 46+ features across 9 modules — consultations, appointments, plans, rates, artisan (karigar) management with gold & wastage tracking. React admin dashboard plus two React Native apps (customer + artisan).
My job-hunt agent once drafted a confidently wrong visa answer. It never shipped — because nothing irreversible happens without a human. That rule is architecture, not policy.
Most rejected work should cost zero model calls. Deterministic hard filters, dedupe hashes and batch prompts come before any token is spent.
With dozens of outcomes, not millions, a Beta-per-bucket model stays honest where a dense one overfits. I can read exactly what my systems have learned.
Request-scoped logs, state-preserving releases, WAL checkpoints before every upgrade. A debuggable system is the difference between an agent and a black box.
Building agents that need to work unsupervised at 3 a.m.? Let's talk shop about reliability, feedback loops, and where the human belongs.
heypremm@gmail.com