- •Runs the Kokoro-82M speech model in the browser on WebGPU via ONNX Runtime — no server, no account, nothing uploaded
- •Highlights each sentence on the real page as it is read, using the CSS Custom Highlight API so the page is never modified
Hi, I'm
Sahil Mahendrakar
Building things. Breaking stuff. Humaning around.
Building in Stealth • Prev. agentic AI at AWS, Columbia '24
About
- →Agentic AI at AWS — building long-term memory for agents on Bedrock AgentCore
- →Former co-founder/CTO — raised $1M, Stanford Blockchain Accelerator
- →Building agent-native tools — for people building alongside AI agents
I’m a product-minded engineer who builds across the stack, from system design to UI/UX. At AWS I work on long-term memory for AI agents; on my own time I build tools for people working alongside them. Lately, I’ve been focused on how humans and agents can build software together.
Projects
Things I've built with AI — from agent tooling to creative side projects.
- •Slack-style channels and DMs where agents are participants — @mention one and it gets to work
- •Persistent agents that do real work: open PRs, run services, and keep their memory across restarts
- •Each agent runs the Claude Agent SDK in its own isolated container with a durable workspace
- •Hook agents up to the tools your team already uses — GitHub, Gmail, Linear, Notion, Slack, Drive
- •Shared kanban board for real-time visibility into every agent's work
- •Isolated git worktrees per agent, with dependencies that auto-unblock as work completes
- •Planning workspace where specs live once and agents read/write them directly
- •Agent-agnostic: drive Claude Code, Codex, and Cursor from one interface
Aristotle
AI-powered e-reader for deeper understanding
- •Spoiler-free contextual Q&A that respects your reading position
- •Inline insights and guided exploration for complex passages
- •RAG pipeline with embeddings and metadata-aware retrieval
- •Move a live session to any machine in your ~/.ssh/config — same session id, context intact, nothing summarized
- •Uncommitted work travels with you and comes back, with same-commit and clean-tree checks so nothing is silently lost
- •Or beam to a cloud sandbox on your own E2B key — no account, no daemon, no hosted service to sign up for
- •Device and folder pickers built from each machine's own Claude Code history, with branch and dirty-file counts
Lyrn Code
AI-powered personalized coding tutor
- •Skills matrix with adaptive practice and real-time evaluation
- •Agent-driven tutoring with function calling and session memory
- •Multi-panel workspace for immersive learning experience
- •Agents search, ask, and share solutions to avoid solving the same problems repeatedly
- •Semantic search surfaces relevant prior solutions from the community
- •Supports web UI, REST API, MCP protocol, CLI, and agent skills
- •Rate albums and build a personal listening profile
- •Agent-based recommendations with spotify integration for album search and metadata
- •Timed Wave sessions where music intensity follows a smooth curve—BPM, density, and prompts evolve with your session
- •Voice steering: natural-language commands classified by Gemini and applied to the live Lyria Realtime stream
- •FocusGuard and app-based routing—music fades when you drift to blocked sites; prompts can follow the app you are in
Writings
Paradee-8M-v1.1
A new decoder head makes Paradee three times faster and removes most of the buzz.
Paradee: Distilling a voice model to one-tenth its size
What I learned shrinking Kokoro-82M into an 8M-parameter voice model that runs on a CPU.
Reflections on Turning 24
When I was fourteen, what did I imagine my life would be like ten years later?
Experience
Columbia University · Bachelor of Science in Computer Science · 4.12/4.0
Software Development Engineer
Sept 2024 – PresentBuilding agentic AI on the Bedrock AgentCore team.
- •Building long-term memory for AI agents on Bedrock AgentCore Memory
- •Previously delivered customer-facing cost optimization experiences (Compute Optimizer, Savings Plans, Cost Optimization Hub)
Agentic AIAWSBedrock AgentCoreTypeScript
Software Development Engineer Intern
May – Aug 2023Built SQL Server license optimization feature for AWS Compute Optimizer.
- •Developed rightsizing and edition downgrade recommendations for SQL Server workloads
- •Enabled customers to reduce license costs by up to 73%
- •Integrated with existing Compute Optimizer infrastructure
JavaAWSSQL Server
Co-founder & CTO
Jun 2022 – Jan 2023IronMill
Led technical development for blockchain security startup.
- •Raised $1M pre-seed funding
- •Selected for Stanford Blockchain Accelerator
- •Awarded U.S. Patent 12,250,316 B2 for zero-knowledge proof system selection
BlockchainCryptographyPython
Cybersecurity Research Intern
Summer 2020 & 2021Westlight AI
Conducted cybersecurity research contributing to federal contracts.
- •Contributed to USAF SBIR Phase I and Phase II awards
- •Co-inventor on U.S. Patent 11,295,029 B1 for filesystem interception and encryption enforcement
CybersecurityResearchPython
Patents
U.S. Patent 12,250,316 B2
Zero-knowledge proof system selection
Method for selecting optimal zero-knowledge proof systems based on application requirements and constraints.
View Patent →U.S. Patent 11,295,029 B1
Filesystem interception and encryption enforcement
System for intercepting filesystem operations and enforcing encryption policies at the kernel level.
View Patent →Now
What I'm currently focused on and interested in.
Current Focus
- →Building long-term memory for AI agents at AWS Bedrock AgentCore
- →Building Fluxx & Context Overflow — tools for people working with coding agents
- →Exploring multi-agent coordination and shared knowledge across agents
Interests
Agent memoryMulti-agent systemsAgent-native toolingHuman–agent collaborationDeveloper experienceSystem design
Contact
Feel free to reach out if you're interested in talking. I'm always happy to connect!
© 2026 Sahil Mahendrakar.