creator of iridium - linkedin agent, the progress app
2yoe @ google via TCS, contributor @ vllm, sglang, ax-platforms

medium

github

Nikhil Kulkarni • Peerlist

<h3>About This Project</h3>This project, fannypack-agents, provides fast, reversible tool calls for LLM agents. It focuses on optimizing the performance of Large Language Model (LLM) agents by shaving off latency and enabling selective undo functionality over an action ledger.Key features include:<ul><li>Latency Shaving: Achieves 1.2–1.8x speed improvements in tool-calling turns by optimizing runtime processes, especially beneficial for slower or repeated calls.</li><li>Selective Undo: Allows for the reversal of specific actions without affecting independent, non-dependent actions. The system clearly indicates whether an action is reversible or irreversible.</li><li>Efficient Tool Integration: Tools are registered using decorators, and their capabilities are defined by their effect classes. The system supports various effect types such as READ_ONLY and IDEMPOTENT_WRITE.</li><li>Reversible Operations: Implements compensator functions to enable the reversal of operations like transfers, ensuring data integrity.</li><li>Snapshotting for Writes: Captures the state before a write operation to facilitate accurate reversals for operations like file edits.</li><li>Prefill and Policy Learning: Optimizes performance by pre-firing read-only calls and learning policies to decline operations that would not yield performance gains.</li><li>Deterministic Routing and Classification: Utilizes deterministic mechanisms for tool selection and classification, reducing reliance on the LLM for these tasks.</li></ul>The project is built with Python and is compatible with Python 3.10+. It integrates with various LLM providers, including Anthropic, OpenAI, and Fireworks, and supports Ollama.Installation can be done via pip or from source. The project also includes extensive documentation and experimental results to demonstrate its performance benefits.

Fannypack Agents

<h3>Iridium provides agent-native access to LinkedIn, enabling AI agents to interact with the platform in real-time on your behalf. User approval required for agentic actions (commenting, messaging, connection requests, fresh LinkedIn data).</h3><ul><li>Live LinkedIn Data: Agents can access up-to-the-minute profiles and posts, ensuring that any content generated or actions taken are based on current information. Data is fetched on demand and expires based on a user-defined retention window.</li><li>Agent Actions: Facilitates a range of LinkedIn activities including commenting on posts, sending connection requests with personalized notes, direct messaging and InMail, replying to conversations, publishing posts in the user's voice, scheduling content, and managing the inbox.</li><li>Human Approval Workflow: Every action initiated by the agent is presented as a draft for review and approval. This ensures that no action is taken on LinkedIn without explicit user consent.</li><li>Clear Logging: An append-only log records every attempted action, including details like the action taken, timestamp, success status, and the exact content sent. This log is accessible within the web app and exposed to the agent for easy auditing.</li><li>Automated Repetitive Tasks: Handles monotonous engagement tasks such as finding relevant people and posts, and drafting original content, allowing users to focus on review and strategic decisions.</li><li>Easy Setup: Designed for quick setup in approximately five minutes, requiring no installation or browser extensions. Integration involves connecting a LinkedIn account via Unipile and adding Iridium as a custom connector in AI models like Claude.</li><li>Flexible Access: Can be used directly within AI conversations via an MCP endpoint or through the Iridium web application, providing users with the flexibility to manage their LinkedIn activities.</li></ul>

Iridium - LinkedIn for AI Agents

Architected a WebSocket-powered AI pipeline with Amazon Bedrock (Nova vision + Llama text) and EventBridge for real-time multi-browser image classification and caption generation, implemented serverless Lambda functions with API Gateway WebSockets, DynamoDB connection tracking, and Datadog observability with distributed tracing and production dashboards.

Bedrock Pipeline

A production-ready ML inference service that transforms informal project descriptions into ATS-optimized resume bullets. Built by finetuning T5-small on 280+ STEM examples, achieving 95% formatting accuracy. Deployed with FastAPI, containerized with Docker, and orchestrated via Kubernetes with GPU acceleration, horizontal autoscaling, and comprehensive monitoring.

Finetuning LLMs

To deploy fullstack app, built 3 tier Kubernetes web app with React, FastAPI, MongoDB, Prometheus and Grafana. Implemented containerized microservices, RESTful API, monitoring dashboards, delivering production ready, cloud app.

Kubernetes Web App with Monitoring

Engineered an AI agent with RAG for intelligent PDF skimming through multi-query workflows, integrated AWS S3 + Lambda with Pinecone vector database and Amazon Bedrock for semantic search, implemented OpenAI GPT-4 function calling for autonomous tool orchestration and S3-triggered ingestion pipeline, delivered document summarization, metadata extraction, and source-cited query responses.

PDF RAG Agent

Deployed a multi-tenant Kubernetes platform serving LLMs with Terraform and EKS. Delivered namespace isolation, RBAC controls, and Prometheus/Grafana monitoring, enabling each tenant to operate in a private namespace with custom system prompts for personalized LLM responses.

Scalable Multi-Tenant Kubernetes for LLMs

Habit tracking app built with Swift and Firebase (Auth + NoSQL DB). Published to the iOS App Store. Helped Nikhil do 180+ days of jogging/walking outside during winter 2024.

Nikhil Kulkarni

creator of iridium - linkedin agent, the progress app 2yoe @ google via TCS, contributor @ vllm, sglang, ax-platforms