Avinash Gosavi
Sepia Dark

Staff Software Engineer · Cisco Meraki · San Francisco Bay Area

Avinash Gosavi

15+ years building production systems at enterprise scale — I care most about distributed systems, API design, and reliability, and enjoy the unglamorous work that makes platforms feel solid: data modeling, integration paths, performance, and being useful when things break. These days that means Rails, React, Kafka, and Kubernetes by day; off hours I build privacy-first AI tools and write about the craft of software engineering.

View Projects Writing

Recently on the site

  • Writing New Paper TL;DR on HNSW, the layered neighbor graph behind most vector search, including the one VaultMem switches to past 1,000 memories. The glossary and VaultMem Part 2 now link to it. →
  • Engineering Search results can now be narrowed to just posts, projects, papers, events or pages (the choice stays in the shareable URL), and you can walk the results with the arrow keys and clear the box with Esc. →
  • Engineering The deploy-time accessibility check now also catches vague link text like "click here", new-tab links missing rel=noopener, and tables without header cells. The "Read more" links on /research now tell screen readers which paper they lead to. →

Full changelog →

Now

Building
Drafting Desk, a writing app where the AI edits and asks questions but never writes the post for you.
Mentoring
Mentoring hackers at Cal Hacks 13.0 in San Francisco, Oct 23–25.
Judging
Judging student projects at HackUMass XIV (UMass Amherst) on Nov 15.
Going to
OSA Con 2026, the open-source analytics & AI conference in San Francisco, on Nov 2.

What else I'm up to →

Hobby Projects

Memory & privacy for AI

VaultMem

Encrypted memory for AI agents

A Python memory library where the platform running the agent cryptographically cannot read what users store. Client-side encryption, encrypted semantic search, PII sanitization. Published on PyPI with a preprint on Zenodo.

  • Python
  • Cryptography
  • AI Agents

Code Live Write-up

Agents & voice

Rena

Voice-first AI health companion

Built for the Gemini Live Agent Challenge 2026 — a health companion you talk to instead of tapping through logging screens. Gemini Live, Veo 2, and Google ADK.

  • Gemini Live
  • Google ADK
  • Voice AI

Code Live Write-up

Memory & privacy for AI

Attractor

Brain-inspired memory engine

A memory engine modeled on Hopfield-style attractor dynamics — retrieval by partial cues and context rather than keys or nearest-neighbour distance.

  • Python
  • Hopfield Networks
  • Memory

Code Write-up

Memory & privacy for AI

daylog

Private voice diary

A voice diary with semantic search where entries and the search index are encrypted at rest, and no content leaves the device to answer a query.

  • Encryption
  • Semantic Search
  • Local-first

Code Live Write-up

All Projects →

Experience

  • Cisco Meraki Staff Software Engineer
    2021 – Present
  • Handshake Software Engineer II
    2020 – 2021
  • CoStar Group Software Engineer II
    2018 – 2020
  • Dominion Enterprises — ForRent Media Solutions Intern → Senior Software Engineer
    2015 – 2018
  • Old Dominion University Teaching & Research Assistant
    2014 – 2015
  • Freewave Tech Solutions Software Developer
    2013 – 2014
  • Indiavidual (via Neosoft Technologies) Ruby on Rails Developer
    2013
  • Allerin Tech Ruby on Rails Developer
    2011 – 2013

Full Experience →

Things I'm curious about

How AI systems should remember — and who gets to read those memories. Brain-inspired retrieval. Running useful models on hardware I own. Keeping big APIs fast. I write up the papers behind my projects in plain language in the Paper TL;DR series — start with six papers that shaped how I think about AI memory. There's also a papershelf and a list of conference talks worth watching.

Recent Writing

All Posts →