Scalable software
with AI systems
judgment built in.

Designed an edge AI agent that can search local context, respond through Telegram, and protect destructive actions behind explicit human review.

Implemented Python services, SQLite/FTS5 retrieval, Linux systemd deployment, and cgroups v2 resource controls for constrained edge compute.

Demonstrated stable AI runtime engineering where inference, search, and automation must share limited CPU, memory, and device resources safely.

Implemented a tiled matrix multiplication kernel and benchmarked throughput against the cuBLAS baseline to understand performance ceilings.

Tuned shared-memory usage, thread-block sizing, arithmetic intensity, memory access patterns, and occupancy while profiling each iteration.

Added low-level engineering evidence for performance-sensitive platforms where implementation details directly affect user-visible scalability.

Created a self-service intake workspace where teams can submit business requests, generate BRD/FRD-style artifacts, route approvals, and prepare Jira-ready ticket payloads.

Mapped request patterns, built tenant auth, client workspaces, automation packs, Jira credential flow, signed webhooks, analytics, and AI research prompts.

Modeled roughly 70% faster clarification on sample requests while preserving human review before work reaches engineering.

Built a workflow portfolio where users can enter process volume, cycle time, errors, cost, integration complexity, and automation readiness.

Implemented a six-dimension scoring model, portfolio analytics, workflow detail pages, webhook subscriptions, settings, and tenant-scoped data APIs.

Modeled $120K+ annualized savings potential on sample data and turned automation ideas into leadership-ready business cases.

Created a clinical queue, intake workflow, record detail view, exception routing, analytics, settings, and research page using synthetic healthcare records.

Used privacy-conscious boundaries, tenant auth, server-side validation, clinical event routes, webhook ingestion, and aggregate-only AI research context.

Modeled roughly 65% faster extraction and 94% exception detection on synthetic data without using real PHI.

Built a merchant workspace for importing product CSVs, validating rule failures, finding duplicates, scoring quality, and managing stewardship stages.

Designed quality rules, batch APIs, dashboard analytics, batch detail updates, merchant settings, webhook subscriptions, and catalog event ingestion.

Modeled 85% duplicate SKU reduction and item setup improvement from days to hours on a 200+ SKU sample catalog.

Created a revenue workspace with CRM import, forecast dashboard, import lifecycle stages, risk detail pages, settings, webhooks, and AI research.

Scored missing next steps, stale activity, close pressure, material deal size, and health signals through server-side APIs and tenant analytics.

Modeled weekly reporting reduction from 4 hours to under 15 minutes and surfaced 23 at-risk relationships in sample data.

Built a 12-stage pipeline for dataset prep, configuration validation, model training setup, evaluation, and experiment tracking.

Used Hugging Face, PEFT/LoRA, Pydantic v2 schemas, config-first orchestration, and automated tests to prevent runtime failure.

Cut time-to-experiment by over 50% and made shared GPU research workflows easier for non-technical collaborators to use.

Built a document-intelligence platform where specialized agents parse, retrieve, and reason over uploaded documents to answer questions with sources.

Orchestrated a multi-agent RAG pipeline — chunking, embeddings, vector search, and reranking — behind a Python backend deployed on Vercel.

Turned static document piles into a queryable knowledge surface, showing end-to-end ownership of a production AI system.

Built a reproducible fine-tuning pipeline covering dataset preparation, training, and evaluation of adapted language models.

Used LoRA/QLoRA parameter-efficient methods, structured configs, and evaluation hooks to keep runs repeatable and inexpensive.

Showed the ability to take base models to domain-specific behavior efficiently on constrained GPU budgets.

Designed a workflow where multiple cooperating agents decompose and complete a task end-to-end.

Implemented agent roles, tool integrations, and orchestration on the Google Agent Development Kit.

Explored modern agent frameworks and coordination patterns beyond single-prompt LLM calls.

Built a research assistant that searches for relevant papers, extracts key findings, and synthesizes them for the user.

Coordinated multiple agents over search, retrieval, and summarization with a RAG backbone.

Compressed literature-review effort — Naga's own publication background applied as a reusable tool.

Built a tool that takes a resume plus a job description and rewrites it into an ATS-compliant, role-targeted resume.

Used LLM-driven analysis to match keywords, restructure content, and score ATS-compliance behind a fast web UI.

Turns a generic resume into a targeted, ATS-friendly application in minutes — AI applied to the job-search problem itself.

Built a platform that helps global citizens compare passports, discover citizenship-by-investment routes, and plan immigration strategies across 190+ countries.

Aggregated country and passport datasets into interactive rankings, pathway explorers, and strategy views behind a responsive web app on Vercel.

Turned scattered immigration data into a single, explorable decision tool for a real-world, high-stakes problem.