I build the systems that let software, and the agents writing it, move fast without moving recklessly: secured pipelines, cheap and well-scoped agents, and automations that talk to real APIs and recover when those APIs misbehave.
Available for automation, SecDevOps and small-business web projects, and for roles where security and automation are the job. Work with me ↓
CompTIA Security+ certified. I treat security as an engineering property, not a review at the end: least privilege by default, secrets kept out of repositories and logs, and untrusted input handled as data, never as code. The capture-the-flag at the bottom of this page is a small, live example: a GitHub Action that processes strangers' input safely.
CI that tests on every platform it claims to support, releases that publish
through short-lived OIDC credentials instead of stored tokens, third-party
actions pinned to exact commits, and builds that fail when they find something
that should not ship. agent-harness runs a 12-job matrix across three
operating systems and four Python versions, and a packaging test scans every
build for leaked private data before release.
| agent-harness | Coding agents get the part of a codebase they need, not all of it: queryable code graphs, character-budgeted context packets, memory that survives between sessions, and human-approval gates. Works with Claude Code, Codex and Gemini. |
| landfirst | When several agents finish work in parallel: did it actually land, and what merges next? Ranks branches by merge order and predicts conflicts. Go, one static binary. |
| compute-pool | Spread workloads across your own machines over Tailscale. Run on the least-loaded one, fan work out, watch it live. Go. |
On token efficiency, honestly: agent-harness is built to cut the tokens
an agent spends, and it already records real input and output tokens from
agent logs. I'm now benchmarking it against the same tasks run without it.
The measured result goes here once it exists, not before.
Most of my work automates the operations of a residential energy-services company in New Jersey. Its CRM has almost no API documentation, so I reverse-engineered the endpoints its own web app calls, and built on them:
- Automated KPI reporting. Technician productivity, pulled from the CRM and payroll and computed per technician, including the numbers that decide bonuses. It runs on a schedule instead of by hand.
- Data pipelines. The CRM's data loaded into Postgres, reconciled against the source, and kept current with idempotent, append-only catch-up loads.
- Data entry, automated. Jobs created back in the CRM with the right program codes, instead of being typed twice.
- Built for a moody API. Backoff when it throttles (HTTP 429), retries that are safe to repeat, scheduled jobs and webhooks.
That's one of many KPI and data-entry workflows I've automated to match how the company actually works. The repositories are private because they hold real customer and employee data.
I maintain my own open-source projects and contribute to other people's. This section keeps itself current: a GitHub Action rebuilds it every day from the GitHub API, so a new project, a release or a newly merged pull request shows up here on its own.
Maintaining · 4 projects
| Project | What it does | Release |
|---|---|---|
| agent-harness | Portable memory & token-efficiency harness for coding agents — queryable code graphs, human-approval gates, and one-command install for Cla… | — |
| compute-pool | Share CPU between your own devices over Tailscale: run commands on the least-loaded device, fan out tasks, live dashboard | v0.1.4 |
| landfirst | Which branch do I merge first? A live merge planner for repos with parallel local work - ranks branches by merge order, predicts conflicts… | v0.4.0 |
| tailnet-sentry | Scan your own tailnet, mark known devices healthy, flag the anomaly, quarantine it. Defensive Go CLI over the Tailscale API. | v0.1.0 |
Contributing · 16 pull requests merged across 3 projects
| Project | What it does | ★ | Merged | Standing |
|---|---|---|---|---|
| affaan-m/ECC | Agent harness for Claude Code, Codex and Cursor. I hardened its safety gate and added the evaluation that measures whether its guardrails change outcomes. | 293k | 14 | #3 of 408 contributors |
| thedotmack/claude-mem | Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More | 107k | 1 | #280 of 300 contributors |
| OpenMathLib/OpenBLAS | OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version. | 7.7k | 1 | — |
| sashs/filebytes | Library to read and edit files in the following formats: Executable and Linking Format (ELF), Portable Executable (PE), MachO and OAT (Android Runtime) | 136 | — · 1 in review | — |
Updated 2026-10-31 · rebuilt daily from the GitHub API by oss/update.py
A flag is hidden in the four tiles above. Each tile flashes one piece of it for a moment, wrapped in a different encoding. Unwrapping it is the easy part: underneath, every piece is encrypted with the cipher a French diplomat got credit for in the 1500s, and each tile's key is the answer to a question about my work. You won't find the answers on this page. You'll find them by reading the projects.
- Cybersecurity. In tailnet-sentry, what tag is an isolated device left
wearing? (the tag's name only, without the
tag:prefix) - SecDevOps. compute-pool's releases after v0.1.0 carry a signed record of where each binary was built. What is it called? (two words)
- Agent infrastructure. landfirst predicts conflicts with a git plumbing command that needs git 2.38 or newer. Which one?
- APIs and automation. When the CRM throttles me, what is it actually saying? (three words)
Keys are the answers in capitals, letters only: no spaces, punctuation or
digits. Each key is at least as long as its piece, so there's no shortcut
around the question. Put the pieces together in order as
FLAG{PIECE1_PIECE2_PIECE3_PIECE4}.
Rules.
- Everything you need is public: this page and my repositories. Nothing requires touching my machines, my tailnet, or anyone else's.
- Solve it, don't attack it. The checker treats every issue as plain data, so tricks just earn "no proof found", and flooding it gets you blocked.
- Submit from your own account. A proof only works for the login that opens the issue.
- Scripts for decoding are fine. Brute-forcing the keys isn't a shortcut: each key is as long as its piece.
- Keep it spoiler-free here. No flag, pieces, keys or walkthroughs in issues or pull requests.
- The flag changes from time to time. Hall-of-fame places are permanent.
Don't post the flag. Issues are public. Post proof that you have it: the SHA-256 of the flag, a colon, and your GitHub login in lowercase.
printf '%s' 'FLAG{...}:your-login' | sha256sum$b = [Text.Encoding]::UTF8.GetBytes('FLAG{...}:your-login'); -join ([Security.Cryptography.SHA256]::HashData($b) | % { $_.ToString('x2') })python3 -c "import hashlib,sys; print(hashlib.sha256(sys.argv[1].encode()).hexdigest())" 'FLAG{...}:your-login'Then submit your proof. A GitHub Action checks it within a minute, replies, and adds you below.
How the checker is hardened
- Visitors submit a per-user hash, never the flag, so a correct answer in a public issue is useless to anyone else. The flag itself only exists in an Actions secret.
- The issue body is read from the event payload by a standard-library Python script and treated purely as data. No untrusted text is ever placed in a shell command or workflow expression.
- The workflow token has no permissions by default, and each job gets only what it needs: verifying can only read, only recording can push, and only replying can touch issues. Git credentials are never left on disk while the checker reads a submission.
- The only third-party action is pinned to a full commit SHA.
- The hall of fame is capped and de-duplicated. Each submission gets its own run, so a burst of issues can't cancel a real one, and recording rebuilds the list on the latest commit and retries, so two solvers can't race. You're only told you made the list once the commit has landed.
- If someone pastes the flag, wrapped or not, in the title or the body, the bot redacts the issue and locks it. The original stays in the issue's edit history until I delete that revision by hand, because GitHub only lets the owner do that.
- The flag pieces are drawn as pixels, not stored as text, so they can't be found by searching the images' source. Even read straight out of the pixels, each is ciphertext under a key as long as itself. Neither the flag nor the keys are in the repository.
Nobody yet. Be the first.
I take on a small number of projects alongside my main work.
| Automations and integrations | Connect the tools a business already pays for: CRMs, job systems, payroll exports, spreadsheets and dashboards. Built for real APIs, with backoff, retries that are safe to repeat, and logs you can read. |
| SecDevOps reviews | A pass over your repositories and CI/CD: actions pinned to commits, least-privilege tokens, OIDC publishing instead of stored secrets, secret scanning and dependency alerts. Findings arrive as pull requests, not a PDF. |
| Websites with security care | Fast sites for small businesses that stay maintained after launch: dependency and framework updates, security headers, and email-spoofing protection (SPF, DKIM, DMARC). |
Write to me with a sentence or two about what's slow, manual or worrying you, and I'll tell you honestly whether I'm the right person for it.
montjeffrey.dev · Portfolio · montjeffrey@gmail.com · New Jersey
Every image here is hand-placed pixel art, animated in pure SVG with no JavaScript and lettered in a hand-drawn 3×5 pixel font. The generator is here.