I'm an AI Engineer who builds full-stack, production-grade systems —
from data architecture to the LLM pipelines that make them
intelligent. My background spans an MSc in Data Science &
Computational Intelligence and a B.Tech in Computer Science &
Engineering, but I learn fastest by shipping real systems rather
than tutorials.
AQLedger, my flagship project, was designed, built, and deployed
solo end-to-end — architecture, backend, database, AI pipeline, and
infrastructure. I care about systems that are genuinely
production-ready: structured LLM outputs over loose prompting,
cost-aware AI pipelines that don't burn money by default, and
infrastructure that's boring in the best way — reliable,
observable, and secure.
Use Cases
Where GenAI meets real backend engineering.
Automated Data Classification
Turn unstructured, free-text input into structured,
schema-validated data using LLMs with strict JSON Schema
outputs — eliminating the inconsistent-value problems loose
prompting causes.
Self-Learning Systems
Historical-lookup caches that get faster and cheaper the more
they're used — falling back to LLM reasoning only for
genuinely new patterns.
Cost-Aware AI Pipelines
User-facing controls over which AI stages actually run — so AI
cost scales with real need instead of blanket automation.
Secure Cloud Infrastructure
Least-privilege identity design and secrets that never touch
source control — backed by Azure Key Vault and scoped service
identities.
Case Studies
Live Project
AQLedger
A self-learning personal finance ledger — every transaction is
entered as free text and automatically classified by an LLM
pipeline, with a historical-lookup cache that gets smarter and
cheaper the more you use it.
Problem
Manually categorizing personal financial transactions was
slow and inconsistent, and no existing tool offered
intelligent, free-text entry tailored to real spending
patterns rather than rigid bank-import categories.
Approach
A cascading classification pipeline — a fast historical
lookup cache checked first, falling back to an LLM with
strict JSON Schema outputs only when genuinely needed.
Thousands of historical transactions were backfilled using a
two-phase taxonomy-discovery-then-locked-classification
process to keep category naming consistent.
Result
A ledger where classification gets both cheaper and more
consistent the more it's used — live-entry AI cost runs at a
few paisa per transaction, with the cache absorbing most
repeat patterns.