// Projects
Selected work
Production agentic-AI systems and intelligent data platforms for CPG and pharma — alongside the engineering and full-stack projects where the craft was built.
CPG · Agentic AI
Agentic Order Management System
Routine business requests rarely fit a single workflow — one conversation might span a question, an update, and an action across systems. I designed a multi-agent system on LangGraph where specialized agents each own a distinct user intent, and a router directs every request to the right one at runtime. It holds context across a conversation, persists session history in a relational database, and runs multi-step workflows through composable graphs and subgraphs. Every LLM call, tool invocation, and prompt is fully traced in LangSmith, a live dashboard tracks application health, and a CI/CD pipeline automates testing, LLM evaluation, and deployment — agentic AI built to be observable, governed, and safe to ship.
Pharma · Enterprise RAG
Permission-Aware Knowledge Assistant
In most enterprises, knowledge lives in SharePoint and Confluence while work happens in Jira, Teams, and Microsoft 365 — so people lose hours context-switching, cross-referencing documents, and re-entering the same information. I built a production multi-agent platform on LangGraph and FastAPI that gives employees one conversational interface to both find knowledge and act on it: retrieve a specification, then raise the Jira ticket that references it, in a single exchange. Specialized agents handle each domain — RAG-based retrieval, Jira, Teams, and M365 — routed at runtime. Security is enterprise-grade throughout: Azure AD / Entra ID SSO, role-based access control, and document-level permission filtering on every retrieval, so each person sees only what they are entitled to. A plug-and-play connector architecture lets each integration be switched on or off independently, with full observability through a self-hosted MLflow tracking server and an Elasticsearch and Kibana dashboard.
Regulated · DevEx
AI Code Review in CI/CD
Pull request reviews are a critical quality gate, but they hinge on senior engineers having time to read every change closely — and under deadline pressure, routine issues quietly crowd out the architectural calls that actually need human judgment. The team was licensed for Copilot but couldn't apply it to Azure DevOps PR reviews, and external review SaaS was off the table because it meant sending proprietary code outside the tenant. I built an AI reviewer that runs as a build-validation step on every PR: an agentic loop reads the change, retrieves the modified files, evaluates them against the team's own standards, and posts advisory, non-blocking comments inline — powered by Claude Sonnet on a Databricks Model Serving endpoint, so no code leaves the tenant and there's no dependency on per-developer seats. A content-addressed cache skips unchanged files across re-runs, idempotent reconciliation keeps the PR thread clean, and the review standards sharpen over time from how leads accept or dismiss comments — freeing reviewers to focus on design while routine issues are caught automatically.
CPG · Self-Serve Analytics
Natural-Language Sales Analytics
CPG commercial teams move at the speed of their data — but brand and category managers usually can't get a number without filing a request and waiting on an analyst, and different teams often define revenue, volume, or share differently. I designed a self-serve analytics layer on Databricks Genie that lets managers ask questions in plain English and get governed, consistent answers in seconds. The work centered on the semantic layer — the ontology, measures, business logic, and context that map everyday business language to the underlying data — so a question like "top brands by Q4 share" resolves correctly and identically for everyone. Curated benchmarking and usage monitoring keep answer quality high as adoption grows, turning a reporting bottleneck into a tool the business uses on its own.
CPG / Pharma · Platform
Self-Healing Data Pipeline
In high-volume CPG and pharma data estates, pipelines break in familiar ways — a supplier quietly renames a column, a feed arrives malformed, a schema drifts — and each break has traditionally meant a failed load, a late-night page, and a manual fix. I designed a self-healing pipeline that pairs a deterministic supervisor with LLM-based repair agents and policy-driven schema governance. It detects and classifies drift, repairs what it can safely resolve on its own, and escalates for human review only when a change would breach policy — all while keeping the Bronze layer immutable so raw data is never silently altered. Pipelines recover without intervention, on-call burden drops sharply, and governance is never traded away for the sake of uptime.
Early Work
Online Compiler IDE
A browser-based IDE that lets users write, compile, and run C, C++, Java, and Python without installing a local toolchain — useful for quick experiments, teaching, and coding practice. The front end captures source and input in the browser while a PHP backend invokes the appropriate compiler on the server, streams back output or compilation errors, and stores submission history in MySQL. Building it meant handling the messy realities of multi-language execution: process isolation, compile-and-run timeouts, and clean error reporting back to the user.
Early Work
Clustering with Genetic Algorithms
An academic study exploring whether evolutionary search can cluster data better than classical methods, using the well-known IRIS dataset as a benchmark. Instead of k-means' fixed iterative refinement, the approach encodes cluster centroids as a genome and evolves them through selection, crossover, and mutation toward a fitness function that rewards tight, well-separated clusters. The project compared convergence behaviour and cluster quality against standard k-means, illustrating where evolutionary techniques help and where their added cost isn't justified.
Early Work
Banking Management System
A desktop banking application that models everyday retail-banking operations — account creation, deposits and withdrawals, balance enquiry, and transaction history — with email alerts on key events. Built in Java with a file- and SQL-backed store, it focuses on the fundamentals that matter in any transactional system: data integrity across operations, clear validation, and a usable interface for non-technical staff. An early, formative project in modelling real business processes as reliable software.
Early Work
GST Management System
A Java application built to handle Goods & Services Tax workflows — recording invoices, computing tax across applicable slabs, and maintaining the records businesses need for compliance and filing. Backed by a SQL database with email-API integration for notifications, it tackled the genuinely fiddly part of tax software: getting calculations right across categories and keeping an auditable trail. A practical exercise in turning regulatory rules into dependable, repeatable logic.
Early Work
Retail Shop Management
A point-of-sale and inventory application for small retail shops, covering daily operations end to end — product and stock management, billing, and basic sales reporting. Built in VB.NET over a SQL backend, it gave shop owners a single place to track inventory levels, generate customer bills, and see how the business was doing day to day, replacing manual registers with something faster and less error-prone.
// no projects match — try a different search
More open-source work lives on GitHub.
github.com/abhishek021 →