Full-stack engineer shipping AI agents, automations, and the systems that keep them honest. Reliability over demos. Fail closed, ground every claim, and never trust a model's own confidence.
Every project below is real. Click any card to see it up close.
A pre-action reliability gate for AI agents. When an agent proposes something that touches real systems, like "approve invoice INV-001 for $12,500," it checks the action against source evidence and answers ALLOW / BLOCK / ESCALATE, with a machine-readable reason the agent can fix against. Deterministic core, fail-closed by design.
View details
A production AI agent that resolves customer refund requests end to end — read the email, verify charges in the ledger, decide against policy, pay once or escalate to a human. Durable runs with idempotent tools, a live approval queue, evals gated in CI, and per-step cost tracing. 97% task completion with zero unsafe payments.
View details
A Chrome extension that summarizes whatever is on screen without leaving the tab. One shortcut turns a long article, PDF, or dashboard into a structured summary with key points, numbers, and action items. Supports six Indian languages, privacy masking, cross-tab comparison, and honest failure when a page cannot be read.
View details
A custom Shopify theme for a plant-based homecare brand, built on Dawn with a bespoke homepage over a fixed animated water scene. Every section is merchant-editable from the theme editor — prices, ratings, badges, and copy all come from the platform, nothing hardcoded. Self-hosted fonts, scroll-reveal motion, and reduced-motion fallbacks.
View details
A flight search engine on the Amadeus API, with live fares, debounced autocomplete, real-time filters, nine currencies, and a four-step booking wizard.
View details
HOS-compliant trip planning for truck drivers. FMCSA hours-of-service law is encoded into every stop, plus DOT-format ELD log sheets exportable to PDF.
View details
A React Native learning platform for Iron Lady, carried solo from first wireframe to launch on the App Store and Google Play, used by many learners.
View details
A retrieval-augmented AI mentor for a board-readiness program. Answers are grounded in the program's own materials, with a deliberate human fallback for sensitive queries.
View details
A lead-management layer on WATI, with real-time pipeline analytics, automated follow-up sequences, and hot-lead alerts.
View details
Multi-stage Zoho Blueprint workflows with 40+ field mappings and self-healing scripts that repair malformed data automatically.
View details
An offer-to-day-one workflow with document collection, auto-generated checklists for every stakeholder, and deadline escalations.
View details
Auto-collected KPIs from CRM, WhatsApp, and email, scored and delivered as role-tailored reports on a schedule.
View details
Attendance-pattern tracking with at-risk-participant flagging and automatic re-engagement before drop-off.
View details
Checklist-based onboarding for sales hires with training modules and manager-facing assessment dashboards.
View details
I'm a full-stack engineer from India. Most of my work lives where software meets consequence: agents that act on real systems, automations a sales team leans on every day, an app many learners carry in their pocket.
The most interesting problem in AI right now isn't making agents capable. It's making them trustworthy enough to leave alone. AgentGate is my running answer to that, and the thinking behind it shows up in everything I build: fail closed, ground every claim in evidence, never trust a model's self-reported confidence.
When I'm not in the editor I write at my blog and build for communities. The best software I've shipped started as someone's daily frustration.
Currently exploring: Kubernetes · Terraform · Grafana
Working on something where an agent, an automation, or a full product actually has to work? I'd love to hear about it. A DM is the fastest way to reach me.