Some Things I Built

A few things I’ve worked on recently: production AI, security systems, Postgres performance, agent workflows, and internal tools that saved people time.

AI / Support

NeetoChat + NeetoKB AI Assist

An AI first responder for support inboxes.

I built a shared AI Assist layer for NeetoChat and NeetoKB so customers can get a useful first answer immediately, even before someone from support jumps in. It reads the question, pulls the right product context from the knowledge base, replies in chat, and hands off to the team when the conversation needs a human.

<10 sec
First response: ~35% less
Queue load: 1 click
Human handoff

AI AssistKnowledge baseSemantic searchSupport inboxHuman handoff

Live

Neeto AI Assist

customer asks

Can I connect my help docs to chat?

incoming question

Customer asks in chat

grounded answer

Uses help articles before replying

human handoff

Support joins when the thread needs judgment

AI replies first

handoff ready

The customer gets a useful first answer right away. If the conversation needs more care, support can take over from there.

Instant first reply

Fewer repetitive tickets

Humans handle the tricky parts

AI / LLM

NeetoBugWatch

A GitHub code reviewer we could actually tune.

I built BigBinary's first production AI app from scratch. It replaced Cursor Bugbot ($15/user/month) with a Gemini + pgvector reviewer that handles around 2,200 PR reviews a month across ~200 repos, usually for $70-$80/month in Gemini spend.

12,000+
Reviews: $70-80/mo
API spend: 195
Repos

Ruby on RailsPostgreSQLpgvectorGemini APIGitHub AppSidekiq

Live

neeto-bug-watch · review #5849

$ bw review --pr 5849
# Specialist ensemble: 4 reviewers + adversarial judge
[generalist]     missing nil guard on user.preferences
[test-semantics] spec asserts << but should be >>
[retry-path]     no backoff on 429 from Gemini
[contract]       deprecated-key still in API response
# Judge: REQUEST_CHANGES (confidence 0.92)
→ posted 4 inline comments, 1 summary

AI / Abuse prevention

NeetoForm Credential Protection

Catches shady password forms before people trust them.

NeetoForm is open enough that anyone can publish a form, which also means abuse protection has to work quietly in the background. I built a safety layer that spots password fields, fake login pages, and phishing-style copy, pauses risky forms, and gives the team a clear review trail.

Auto pause
Fake login pages: <1 min
Review time: Keep moving
Clean forms

Abuse detectionAI classificationForm scanningAccount safetyTeam alerts

Live

Form safety layer

risky form

Looks like a password collection page.

paused

password traps

Forms that ask for secrets they should not collect

fake login pages

Copy and layout that pretend to be another product

phishing language

Urgent prompts designed to rush people into mistakes

Good forms keep moving. Suspicious ones pause for review before they can collect anything sensitive.

Performance

NeetoCal Performance

Postgres work on the parts users were waiting on.

Worked through the slow paths in NeetoCal, which handles millions of events and bookings. Cut some query costs by up to 99.95%, removed 30s timeouts from customer-facing slot APIs, and cleaned up 28.6 million rows left behind by a Google Calendar sync runaway.

99.95%
Cost reduction: 89%
Latency cut: 28.6M rows
Data leak fix

PostgreSQLGIST + tsrangeSidekiqpganalyzeDatadog

Live Read more

Query cost · before / after

before281,139

after131

99.95% reduction

Infrastructure

NeetoDevBot

Claude Code agents running inside Kubernetes jobs.

Built the Dependabot flow around a Rails 8 + Kubernetes control plane. It sends Claude Code agents into one-shot pods, lets them work stuck Dependabot PRs, and usually gets the fix shipped in under 10 minutes.

100%
PR merge rate: 3
Repos coordinated: <10 min
TTR

Rails 8PostgreSQLSidekiqKubernetesClaude CodeOctokit

control plane topology

Tooling

Neeto Claude Skills

A shared skill library for engineers and agents.

Put together reusable prompt and skill bundles for Claude Code. A lot of the work was making the loops less fragile: race-condition fixes, false-positive checks, status polling, and one source of truth that works from a terminal or an autonomous Kubernetes pod.

13
Human skills: 26
Agent skills

Claude CodeMarkdown skillsBashGitHub Actions

skills for humans and agents

github shelf

Some of my personal projects and contributions.

personal repos

Claude Code workflow

ccsesh

yedhink/ccsesh

A tiny Claude Code helper that fuzzy-searches local sessions and resumes the one I pick.

Shell3 stars0 forks

Vim, tmux, macOS

dotfiles

yedhink/dotfiles_ikigai

My actual shell/editor/tmux setup. The repo where workflow taste keeps leaking into code.

Shell148 stars16 forks

minimal git prompt

bashed-on-a-feeling

yedhink/bashed-on-a-feeling

A fast minimal git prompt written in bash. Old, scrappy, and still very much my kind of thing.

Shell130 stars9 forks

upstream trail

PR #54649 merged

Ruby on Rails

rails/rails

Merged docs PR around Rails generator behavior and table-name assumptions.

Ruby58.4k stars

Open PR

docs upstream

React docs

reactjs/react.dev

React documentation codebase. One of the upstreams I keep close because product UI and docs shape how people learn.

JavaScript11.7k stars

editor upstream

Neovim

neovim/neovim

The editor upstream that has influenced more of my daily workflow than almost anything else.

Vim Script99.3k stars