BPAN Automation Web App

A production automation tool built and shipped to production in one week.

What it does

• Generates unique 21-digit battery identifiers (BPAN) for every unit that comes off the production line
• Attaches a QR code to each identifier that links to the unit's technical data sheet
• Manages the lifecycle of these identifiers — generation, lookup, audit, and bulk operations
• Replaces a manual process that was slow and error-prone

Why it matters

• Every battery Hykon produces gets a BPAN — it is the canonical identifier used across operations, sales, and after-sales
• The QR-linked data sheet means anyone in the field can pull up a unit's full spec sheet by scanning the code
• Built and shipped to production in one week, after a deeper scoping process had stalled for months

Highlights

• End-to-end ownership — schema design, UI, generation logic, and deployment
• Designed for the floor: large touch targets, fast input flow, minimal clicks per BPAN
• Hand-off ready — the team continues to maintain and extend it post-internship

Hykon Automation Hub

A self-proposed internal platform — built because the team needed one place to find every AI tool that was being built across the company.

What it does

• Centralizes every AI tool built by the team into a single searchable hub
• Each tool has a description, owner, status, and access link
• Reduces the "I didn't know we had a tool for that" problem that grew as the team shipped more automations

Why it matters

• Self-proposed — not a top-down ask. Pitched to leadership based on observed friction
• Recognized as a key company initiative; an ERP-integrated Phase 2 is underway
• A real example of seeing a structural problem and building the fix, not waiting for a ticket

Highlights

• Discovery problem framing — defined the user (every non-technical team member) and the friction (no central catalog)
• Lightweight MVP that earned leadership buy-in for a Phase 2
• Phase 2 includes direct ERP integration, which means the hub becomes a workflow surface, not just a directory

Naukri AI Candidate Shortlisting Tool

A two-agent Chrome extension that screens and ranks job candidates entirely client-side.

What it does

• Lives as a Chrome extension on top of the Naukri recruiter dashboard
• Two collaborating agents: one screens the candidate profile against the job description, the other ranks and explains the result
• Runs entirely client-side — no candidate data leaves the browser

Why it matters

• Cuts manual HR screening time by ~54 hours per month
• Privacy-preserving by design: the only model calls are made with the user's own API key, and no candidate data is logged server-side
• Two-agent split gives a clearer separation of concerns than a single mega-prompt — easier to debug, easier to extend

Highlights

• Client-side LLM orchestration
• Agent prompts and tool definitions tuned for the Naukri profile schema
• Worked end-to-end on real recruiter workflows

AI Candidate Scoring — Zoho Recruit

An in-app Deluge function that auto-scores and summarizes candidates inside Zoho Recruit.

What it does

• Runs as a custom function inside Zoho Recruit — no separate app, no context switching
• Reads a candidate's profile and the job they're being considered for
• Auto-scores the candidate against the job and produces a short, recruiter-friendly summary

Why it matters

• Speeds up recruiter review by surfacing the signal that matters
• Lives where the recruiters already work, which is the difference between "AI tool" and "AI tool that actually gets used"

Highlights

• Native Deluge integration with the Zoho Recruit data model
• Output format tuned for the recruiter reading flow — score, then summary, then the strongest signal

Evocode: AI-Powered Learning Ecosystem

A personal project — an adaptive learning platform that uses LangGraph-orchestrated LLM agents to generate personalized syllabi and mastery-based assessments.

What it does

• Builds an adaptive learning path for a learner, based on their stated goal and prior knowledge
• Uses a LangGraph-orchestrated set of LLM agents to generate the syllabus, the lessons, and the assessments
• Re-tests concepts the learner gets wrong, escalating to harder variants as they demonstrate mastery

Why it matters

• A real testbed for agent design — multi-agent flows are easy to draw, hard to make robust, and Evocode is where I push on that
• Useful in its own right: the assessment quality on harder topics is better than most static quiz generators I've seen

Highlights

• LangGraph agent orchestration with stateful learner profiles
• FastAPI backend, Google Firestore for learner data
• React frontend with progress dashboards
• Mastery loop: a concept is "done" only after the learner passes a generated assessment

If you build anything with LLMs, you'll spend a lot of time wrapping them in an API. I've tried Flask, Django REST, and a few others — but FastAPI is the one I keep coming back to.

It's simple. The decorator-based routing feels natural, the type hints double as request validation, and the auto-generated docs at /docs mean I never have to write a separate API spec. For an AI engineer who wants to ship agents and RAG pipelines fast, that combination is hard to beat.

It also plays nicely with the rest of the AI stack. Async support, Pydantic models for structured LLM outputs, easy streaming responses — everything I need for an LLM-powered endpoint is one import away.

It's not the right tool for every job, but for the kind of AI apps I like to build, it's the closest thing to a perfect match I've found.