Software architect with 11 years across TypeScript/Node.js, Python, Go and AWS. I build cloud and release infrastructure, web and mobile products, data systems, and AI agent tooling with hard safety boundaries.
Currently a software architect at a drone-operations company.
Open to senior engineering and architecture roles (remote).
Past the line, the statement runs against the SQLite database.
Held: writes need a human to approve this exact statement.
Demo of how sqlguard-mcp gates writes. In the real tool the approval happens out of band with sqlguard approve <token>.
Transaction
A batch of writes
Past the line, the transaction is applied.
Held: a commit is applied only after it is written to the write-ahead log.
Demo of how atlas-graph-db commits. Its engine appends each transaction to a write-ahead log, synced to disk by default, before applying it, and recovers by replaying the log.
Release
runon 0.14.0, ready to publish to PyPI
Past the line, the release is published to PyPI.
Held: a release publishes only after its tests pass.
Demo of runon's release workflow: it runs the tests, then publishes to PyPI through Trusted Publishing.
Test run
A new run, compared with the baseline
Past the line, the result is reported.
Held: a difference counts as a regression only if it clears the run-to-run noise.
Demo of how noisefloor judges a change: it runs your target several times, measures the spread, and reports each case as regressed, broke, improved or unchanged.
Candidate
A candidate article from your sources
Past the line, it can be drafted in your editorial voice.
Held: it is drafted only if it clears the quality threshold, adjusted for the model your workspace picked.
Demo of My Stream's score gate. Try the whole pipeline in the Products section.
Question
How do I submit an expense report for travel?
Past the line, the model answers with citations.
Held: it answers only if retrieved evidence clears its confidence floor.
From rag-knowledge-agent's README: a question its sample documents can't answer is refused before any model call.
Proposed by the triage agents
scale_up on checkout-api
Past the line, the action runs through a dry-run adapter.
Held: the remediation waits for a human to approve, edit or reject it.
Demo of triage-graph's approval gate. It runs on fixture incidents with a dry-run adapter.
Five areas, and the work that shows each one.
Principles I come back to, each quoted from the code or docs where it lives.
“This module, not any prompt, decides what may run.”
“There is no tool that approves.”
“runs your target N times, measures the spread, and only calls something a regression when the delta clears it.”
“tells the model what was wrong, not just that it was wrong”
“the confidence gate spends no tokens”
“Adding a capability means adding a directory. The tool never changes.”
Thirty-two public repos with 3,292 passing tests as of October 2026, one dot per test. After three featured repos, the route flies five legs that group the rest by what they do.
The three ringed in magenta come first: two agent tools that stop for human approval, and a graph database built from scratch. I ran every suite between 13 and 16 October 2026, and each count matches the repo's CI. Where some tests need a service such as PostgreSQL, Redis or an Android emulator, the panel says so, and where this machine lacks something, such as ROS 2, the panel says that too.
AI systems
134 passing tests
Multi-agent incident triage on LangGraph with a human approval gate enforced in code, plus a CrewAI port and a written comparison of the two.
uv run pytest (134 passed, 1 skipped)AI systems
139 passing tests
An MCP server that lets an agent query a SQLite database and refuses any write until a human approves that exact statement.
sqlguard approve <token> out of band. Each approval is single-use and bound to the exact statement. SQLite only.go test -race -count=1 -json ./... (65 tests plus 74 subtests)Data systems
686 passing tests
A graph database written from scratch in TypeScript, with a write-ahead log, crash recovery, atomic transactions and a Cypher-like query language (AQL).
pnpm test (468 engine and server tests plus 218 web app tests)591 passing tests in six repos
140 tests
Lab instruments for agents, behind interlocks
78 tests
Replays an agent's tool calls on a new model
65 tests
Audits MCP servers in a traced sandbox
109 tests
Cited answers, or a refusal
105 tests
Schema-checked routing across LLM providers
94 tests
Network and SSH tools, denied until opted in
202 passing tests in six repos
39 tests
Local-first notes on a fuzzed CRDT
38 tests
Raft under a deterministic fault simulator
35 tests
Every state a power cut can leave on disk
40 tests
What a PostgreSQL migration will lock
29 tests
Breaking-change checks for JSON events
21 tests
Time-series compression, best codec per block
Latency, limits, alerts and flaky tests, measured instead of assumed.
Open the group page499 passing tests in seven repos
49 tests
Request time budgets across HTTP services
33 tests
Rate limiters under clock skew and restarts
37 tests
Before-and-after OpenTelemetry trace diffs
71 tests
Alert rules replayed against past incidents
202 tests
Regression tests for noisy systems
47 tests
Why each flaky test flakes
60 tests
A real-time dashboard, measured in the browser
523 passing tests in four repos
36 tests
What an npm upgrade can newly do
38 tests
Admission policies replayed on the audit log
27 tests
Who gains or loses access in a schema change
422 tests
Shell programs across servers over SSH
Flight safety, firmware and geology, tested in simulators and on emulators.
Open the group page518 passing tests in six repos
96 tests
Forwards only setpoints proven inside the envelope
237 tests
Proves a mission clears the terrain
67 tests
CI check for PX4 and ArduPilot parameters
33 tests
A/B firmware updates that survive power cuts
46 tests
Borehole logs to a ground model with intervals
39 tests
Offline Android notebook for boreholes
Products I ship, mostly solo: a web app for AI publishing workflows, a Quran reader on the web, and a long-running Quran app for Android.
feed.geohashim.com
Web app
An AI publishing platform. Scrape sources, score them against a quality threshold, and draft articles in your editorial voice. Workspaces and scheduled automation.
It's an Angular and Ionic web app, with its backend on AWS: Amplify, Lambda, an AppSync GraphQL API, Cognito sign-in and Stripe billing.
Built in Anthropic runs out of the box, the rest with your own keys.
Effective threshold
= clamp(60 + 0, 20, 95)
effective = clamp(platform threshold + model offset, 20, 95)
Cheaper models score lower on the same rubric, so My Stream adjusts the threshold per model and they aren't silently rejected.
Turn it off to see the problem it solves: a cheaper model scores lower on the same rubric and most items get rejected.
Run type: on demand
A run started from the article queue is the only path through the score gate. Scheduled runs, daily or on weekdays at 08:00 UTC, write from the workspace's topic list and skip scraping, scoring and the threshold.
Simulated run with made-up items. Each dot is an item with a random base score, shifted by the model's offset: the higher it flies, the higher it scored. Items below the magenta line peel off at the Score gate.the further right it flies, the higher it scored. Items left of the magenta line peel off at the Score gate. The starting threshold of 60 is an illustrative setting. To show more traffic, the demo scores every item and drafts the top 3 that pass; the product's platform settings default to 5 scored and 1 drafted per run. The stages and the threshold formula are the real ones.
Sources
You point it at RSS feeds and topics, per workspace.
Scrape
A Lambda fetches each source and queues candidate items.
Score
The model you pick rates each item, up to a per-run limit, against a fixed rubric and a quality threshold, adjusted per model so cheaper models aren't silently rejected.
Draft
The highest-scoring items that pass, up to a per-run limit, are drafted in your editorial voice: voice, reading level, word count, keywords and citation rules.
Publish
You review the queue, edit if you want, and publish.
Example paragraph in this voice
Houseplants need less water and more light as winter sets in. Moving pots closer to a bright window, while keeping leaves clear of cold glass, helps most common varieties through the darker months. Watering can wait until the top of the soil feels dry.
Your plants are probably a bit sleepy right now, and that's fine. Give them the brightest window you've got, keep them off the cold glass, and go easy on the watering. They'll perk up again once spring rolls around.
During the winter months, houseplants require less frequent watering and greater access to natural light. Pots should be positioned near a bright window, away from direct contact with cold glass, and the soil should be allowed to dry between waterings.
Shorter days in winter slow the growth of most houseplants, which in turn lowers their water requirements. Care routines should therefore prioritize light over watering frequency: plants placed near bright windows, but away from cold glass, are less likely to suffer leaf damage during this period.
So, your plants look a little tired lately? That's normal for this time of year. With shorter days they grow slowly, so they need less water than you'd think. Move them somewhere bright, keep them away from the cold glass, and check the soil before you water.
This simulated run
In the product, the page polls the job record for run progress every 1.5 seconds.
Works with 10 AI providers: Anthropic, OpenAI, Google, xAI, DeepSeek, Mistral, Groq, Cerebras, OpenRouter and SambaNova. Anthropic runs out of the box, the rest with your own keys.
mushaf.geohashim.com
Web app
A focused, ad-free digital Quran reader for the web, built for fast page navigation, comfortable reading typography and a layout that respects the printed Mushaf.
On Google Play
Android app
My long-running Quran app for Android, with reciter audio, bookmarks and offline reading. It has been in users' hands for years.
Get it on Google Play How it works: My Stream - القرآن الكريم
I'm a software architect at a drone-operations company, with 11 years across TypeScript/Node.js, Python, Go and AWS.
I build cloud and release infrastructure, web and mobile products, data systems, and AI agent tooling with hard safety boundaries. On this page, that means an AWS backend with sign-in and billing, a release workflow that runs the tests before it publishes, web apps and an Android app, a graph database written from scratch, and agent tools that stop for human approval.
I also ship my own products, mostly solo: applied AI for publishing workflows, and a long-running line of work in Islamic resources, including a digital Mushaf and a Quran Android app that have been in users' hands for years.