Projects

Production-oriented systems, autonomous AI agents, and full-stack platforms built around real engineering challenges.

Vector ~68ms · Chat ~1.1s (Benchmark)

Karya

Offline-first multilingual AI learning platform.

An offline AI education platform combining quantized on-device LLM inference, multilingual RAG across 22 Indian languages + English, document OCR, and interactive whiteboard tools.

Architecture Highlight

Implemented an offline-first AI architecture using a quantized 1.5B LLM designed for 4–8GB RAM environments.

Key Engineering Deliverables

•Local LLM tutoring: Quantized DeepSeek-R1-Distill-Qwen-1.5B designed and tested for 4–8GB RAM environments
•Multilingual RAG & NMT: Local semantic retrieval with ChromaDB and separate NLLB-200 translation across 22 Indian languages + English
•Automated study materials: Summaries, notes, flashcards, MCQs & diagrams

ReactNode.jsMongoDBChromaDBllama.cpp

RAPTOR Tree · Async Queue

AuraNow

AI social media intelligence transforming video comments into hierarchical RAPTOR topic trees.

An AI-powered video comment intelligence platform that automatically clusters YouTube comments into interactive RAPTOR topic trees, audience sentiment analytics, toxicity audits, and conversational insights.

Architecture Highlight

Engineered an asynchronous NLP pipeline combining vector embeddings, UMAP dimensionality reduction, HDBSCAN clustering, medoid selection, and recursive RAPTOR tree construction.

Key Engineering Deliverables

•Hierarchical RAPTOR Topic Trees: Clusters comments into high-level themes and sub-topics using representative medoids, quotes, and keyword extraction
•Representative Medoids: Uses real comments as cluster representatives for interpretable topic summaries instead of synthetic centroids
•Audience & Sentiment Insights: Engagement velocity tracking, sentiment distribution, engagement breakdown, and emoji/word clouds

FastAPIPythonReact 18RAPTORUMAP / HDBSCANMongoDB

GitHub

Fullstack

7 Languages · Real-time Sync

Code with Buddy

Real-time collaborative code editor with live execution.

A browser-based pair programming tool featuring real-time collaborative code editing synchronized via WebSockets, multi-language execution across 7 languages, and integrated chat.

Architecture Highlight

Engineered low-latency document sync by filtering update events and using targeted handshakes for late-joiners.

Key Engineering Deliverables

•WebSocket-based real-time editing: Low-latency collaborative document synchronization
•Late-joiner state sync: Fast one-time handshake without replaying full edit history
•Multi-language execution: Remote code compilation across 7 languages via Judge0 API

ReactNode.jsSocket.IOExpressCodeMirror

GitHub

Fullstack

29/29 tests passing

IncidentFlow

Mission-critical incident management platform.

A full-stack incident management system with automated SLA enforcement, Celery asynchronous alerts, RBAC, audit trails, and reliability analytics.

Architecture Highlight

Engineered an idempotent asynchronous SLA engine that reliably notifies on approaching deadlines with an 80% duration warning threshold.

Key Engineering Deliverables

•Automated SLA deadline monitoring: Idempotent Celery engine with 80% duration early-warning alerts
•Role-based access control & audit trail: Immutable logging of every status and ownership change
•Mandatory postmortem gating: Enforces completed root-cause analysis before closing incidents

PythonFastAPIReactPostgreSQLCelery

Agents

Verify-First Workflow

Terminal Agent

Verify-first autonomous software engineering agent.

An autonomous AI coding agent that plans, modifies, tests, and independently verifies real-world codebases with sandboxed isolation and automatic rollback.

Architecture Highlight

Implements a verify-first workflow by executing an independent test runner in an isolated sandbox for each code modification.

Key Engineering Deliverables

•Autonomous agent loop: LLM planning, controlled CLI execution & test verification
•Independent verification: Separates code generation from test execution and verification inside an isolated sandbox
•Automatic rollback & recovery: Classifies failures across 12 categories with SQLite checkpoints

PythonDockerSQLiteGitPytest

Agents

40+ benchmark tasks

Agent Bench

Autonomous AI agent benchmarking framework.

A CLI toolkit for benchmarking AI coding agents across 40+ programming tasks using deterministic rule-based scoring and full execution trace logging.

Architecture Highlight

Automated multi-task benchmarking of AI agents with deterministic rule-based scoring for fair, non-stochastic comparison.

Key Engineering Deliverables

•40+ structured benchmark tasks: Real-world programming and tool-use scenarios
•Deterministic rule-based scoring: Avoids stochastic LLM judging through deterministic rule-based evaluation
•Execution trace logging: Complete JSON audit trail with tool calls, timestamps, and timing

PythonCLIYAMLDockerJSON

Agents

Adaptive Routing · Cost Optimization

Model Router

Adaptive LLM selection proxy designed to optimize inference costs through intelligent routing.

An intelligent proxy gateway that dynamically routes LLM queries based on prompt complexity—directing simpler queries to lower-cost models while reserving premium models for complex tasks.

Architecture Highlight

Engineered an adaptive LLM gateway designed to reduce API costs by dynamically routing prompts based on query complexity.

Key Engineering Deliverables

•Prompt-based model selection: Analyzes complexity to route between local models and flagship APIs
•OpenAI / Claude API proxy: API-compatible gateway and proxy for standard client SDKs
•Cost-efficient routing: Routes simpler queries to lower-cost models while reserving premium models for complex tasks

PythonFastAPIOpenAI APIAnthropic APILocal LLMs