About Me
My approach to AI engineering has evolved from writing code that works to designing systems that continue to work under real-world constraints. Through hands-on experience building production-facing AI pipelines and backend services, I’ve learned to prioritize reliability, observability, and controlled failure over one-off solutions.
When working on a system, I consistently think through:
- What is the most likely way this can fail?
- How would we detect and diagnose it in production?
- What is the simplest rollback or recovery path?
My current work spans agentic AI workflows, backend API development (FastAPI), and applied machine learning, with a focus on connecting ML components to reliable, well-structured services that can be used in real applications.
My Skills
Core Languages & Backend
Python
FastAPI
REST API Design
SQL
LLMs & Agentic Systems
Agentic Workflows (LangGraph)
Retrieval-Augmented Generation (RAG)
LangChain
LLM APIs (Gemini, OpenAI, Groq)
Prompt Engineering
RAG Evaluation (Ragas, TruLens)
Retrieval & Vector Systems
Sentence-Transformers
Vector Databases (ChromaDB / pgvector)
Embedding Models
Machine Learning & Deep Learning
Supervised Learning
Feature Engineering
Transformer Architectures
Fine-Tuning
PyTorch
ML and Data Tooling
NumPy
Pandas
Scikit-learn
Infrastructure & Deployment
Docker
AWS
pgvector
PostgreSQL
API Design
Observability
My Projects
Agent Breaker – AI Agent Adversarial Testing Framework (PyPI) - 2,300+ Downloads
• Built a pip-installable framework to test LLM agents against prompt injection, goal hijacking, and unsafe tool usage.
• Developed runtime introspection + targeted attack generation for agent-specific exploits; plug-and-play with LangGraph systems.
• Engineered a hybrid evaluation system combining rule-based logic with a PyTorch ML classifier (97.8% accuracy), reducing false positives from 30% to <5%.
• Achieved 2,300+ downloads (500+ real) with 140+ weekly installs, demonstrating adoption among AI agent developers.
Tech Used:
Python
Pydantic
Typer
LangGraph
PyTorch
Not-Jarvis - Stateful AI Agent Framework
Built a stateful, multi-turn AI agent using LangGraph with an iterative single-step planning loop, executing one action per iteration with explicit completion checks.
• Designed a hybrid Python + LLM architecture where Python performs deterministic URL extraction and normalization, achieving zero URL hallucination and duplicate action prevention.
• Implemented persistent conversation memory (PostgreSQL checkpointer), Server-Sent Events (SSE) streaming, and schema-validated planning (Pydantic) to ensure reliable, debuggable agent execution.
Tech Used:
Python
FastAPI
LangGraph
LangChain
Gemini API
PostgreSQL
Pydantic
VidQuery - Low-Latency RAG System for Long-Form Video QA
Designed and built a sub-5s latency RAG system for long-form YouTube videos, implementing URL-based transcript caching to reduce compute costs and latency for repeated queries.
Implemented hybrid dense + keyword search with multi-query retrieval while enforcing strict top-k and token caps to keep LLM context small and cost-efficient.
Tech Used:
Python
LangChain
Gemini
Hugging Face
Sentence-Transformers
ChromaDB
Blaze - AI Web Page Summarizer and ChatBot - Chrome Extension
A multi-functional AI browser extension providing instant summarization and enabling users to ask in-depth questions about any webpage’s content.
This tool leverages a custom-built RAG pipeline and a Python/Flask API to deliver precise, AI-driven insights directly in the browser. Implemented Map-Reduce for efficient Web-Page summarization, and ChromaDB for Vector Store.
Tech Used:
Python
LangChain
Google Gemini
Hugging Face
ChromaDB
JavaScript
My Blog
I write about Machine Learning, AI, and the projects I'm working on. Check out my articles on Medium.
Read on MediumProfessional Experience
Founding AI Engineer
Jetty AI (Remote, India) | Mar 2026 - Present
- Built the core orchestration engine behind OpenJetty, an AI crew for local businesses that converts plain-English intent into a execution plan (reply, actions[]) through a single LLM routing call, dispatching 25+ action types across 5 specialized agents using live business and market data.
- Engineered the multi-stage pipeline around it — deterministic triage (DB-backed metric answers, greeting fast-paths, and multi-turn conversational state machines) keeps most requests off the LLM, so only genuine reasoning reaches the planner before plans execute server-side (model proposes, code executes).
- Designed a server-enforced human-in-the-loop workflow for bulk and customer-facing actions, with persistent approval queues; built durable multi-goal state in Postgres so concurrent goals remain isolated across turns.
- Led architecture and code reviews across Jetty AI products, with improvements adopted into production; evaluated engineering candidates for the founding team.
Quant AI Engineer Intern
Fidura AI (Remote, India) | Oct 2025 - Jan 2026
- Designed and owned a production-grade RAG extraction pipeline converting 500+ page tender PDFs into 19 schema-validated fields, replacing multi-day manual review workflows through hybrid retrieval (sentence-transformers + pgvector), Docling ingestion, bounded-concurrency LLM execution, and page-level provenance for traceability.
- Built an agent-orchestrated supply-chain intelligence platform that dynamically constructed origin-to-destination supply-chain knowledge graphs (manufacturers, processors, logistics providers, regulators, and suppliers), enriched each node with source-backed news intelligence, and supported natural-language querying for supply-chain planning and risk analysis.
- Integrated 3+ custom financial tools, including strategy backtesting, simulated order execution, and hedge-ratio calculation, to support quantitative risk management workflows.
- Architected and implemented a LangGraph-based agent orchestration system to automate quantitative workflows from research to simulated execution.
- Built and maintained multiple FastAPI services (20+ endpoints) exposing AI agents and financial tools as external APIs, handling request validation, concurrency, and downstream integrations.
Head Of Operations (Technical Internship, Full Time)
Young Compete (Remote, India) | Jul 2024 - Jan 2025
- Developed 3+ tech projects, including Prof Connect, enabling streamlined networking for 1,500+ students and teachers.
- Automated data scraping from 5 sources for the off-campus placements page, utilizing APIs, and improved access rate by 30%.
- Led a 10-member team to deliver technical solutions and organized events, focusing on user engagement and community building.
Achievements & Activities
- Qualified for Smart India Hackathon (SIH) 2025 (Internal Round)
- Serving as Co-Lead of Machine Learning and Artificial Intelligence at GitHub Community GITAM.
- Organized 6+ ”GAAC-MasterClasses” for Gitam Aero Astro Club (affiliated with MIT, Boston), achieving 60+ attendees per session, focusing on technical education and aerospace topics.
- Competed in national-level events like IIT Kanpur’s Techkriti (Maneuver Bot, Astro Quiz) and ISRO’s Bharatiya Antariksh Hackathon (Lunar Crater Locator), among 50+ teams.
- Secured 2 nd place in the hackathon CODEX (for my SkillCheck Project) conducted by Young Compete.
- Member of Art of Living’s YES+ Group since 1st year of college.
Certifications
- Building with the Claude API (MCPs, Agents, Prompt Evaluations) - Anthropic
- Fine-Tuning Language Models with Hugging Face - Hugging Face
- Introduction To Fine-Tuning - The LLM Course - Hugging Face
- Introduction to Transformer-Based Natural Language Processing - Nvidia
- Anthropic - AI Fluency : Frameworks & Foundations - Anthropic
- IBM – Databases and SQL for Data Science - IBM
- Google AI Essentials – Coursera - Google
- AI for Everyone – deeplearning.ai - deeplearning.ai
My Resume
Interested in learning more about my qualifications? Download my resume for a detailed overview.
Download ResumeGet In Touch
Have a question, a project idea, or just want to connect? Feel free to reach out!