Forward Deployed Engineer · portfolio

Armando Gomez

Integrations that fail safely in someone else's environment: who is allowed to call what, what happens when a source dies, and what a system does when the data it reads is written by an attacker. Every figure on this page is produced by a command, not typed in.

Test suite — loading
Tool surface — loading
Side effects — loading
Portfolio 2 / 5 built 2 write-ups published, 1 project not started

Who can call what

Pick a role — the server really does register only these tools for it

Tool surface for the selected role
ToolRequires scopeStatusWhat it does

A tool the role doesn't grant is never registered, so it isn't in the list the model chooses from, and a direct call fails as an unknown tool. Scopes come from config/policy.toml; an unknown role gets nothing at all.

When the data is the attack

A pull request title is written by a stranger and lands in the model's context

As filed by the attacker hidden characters shown
loading…
As this server returns it labelled untrusted
loading…
Characters removed
The instruction is still readable — on purpose.

Quietly rewriting third-party text would hide the attack from the human reading the output, and phrase-matching for jailbreaks misses rewordings while mangling legitimate bug reports. What's removed is only what hides: zero-width, bidi-override and ANSI characters. What stops the attack is the previous section — as viewer, the tool this payload names was never registered, so there is nothing to call. Telegram stays dry-run until a human sets TELEGRAM_DRY_RUN=false.

What a client gets when something breaks

Real envelopes, produced by running the real guard

loading…
—

Why an envelope instead of an exception

A model can't act on a stack trace, and a crash in one source shouldn't take the others with it. Every tool answers with the same shape, so the client can tell "wait 42 seconds" apart from "this will never work" without parsing prose.

Unexpected exceptions become internal with a generic message: exception text can carry tokens, so detail stays in the server log.

Test suites

Portfolio track

Each project moves through the same five stages

MCP server: GitHub, SQLite, Telegram

Role-scoped tools, per-source degradation, alerts that survive an outage, hostile input treated as data.

fde-mcp · Python · mcp 2.2

Three connectors, outbox fallback, injection defences, runnable demo. ADR pending.

Documenting

Evaluation framework for a real use case

37 cases over reads, writes, refusals, ambiguity, injection and error handling — scored for safety as well as accuracy.

fde-evals · results committed · 45.9% -> 94.6%, 8 unsafe -> 0

Two agents scored, results and known gaps published.

Publishing

Enterprise integration with real constraints

Legacy system or inconsistent data; document what broke, the undocumented behaviour found, and the stakeholder impact.

System not chosen yet

Not started

Queued

Architecture Decision Record

Two written: binding authorisation to the process, and bounding what injected text can reach instead of detecting it.

docs/adr/0001, 0002

Rejected options, costs, and what would reopen each decision.

Published

Production post-mortem

The demo crashed on every console but mine: an encoding assumption no test could catch, and the habit that found it.

docs/postmortems/2026-09-15-demo-crash-on-legacy-console.md

Timeline, wrong turns, root cause, and three prevention changes.

Published

What a reviewer checks in 15 seconds

Scored against this repo today — every line backed by a command

README leads with the problem and lists limitations

Met

Problem, why these sources, auth per source, threat model, degradation table, honest limitations.

Secrets stay out of the repo

Met

Every tracked file is scanned for credential shapes, and the patterns are self-tested against realistic values so the scan can't pass vacuously.

What happens when things go wrong

Met

Every documented failure maps to a stable kind with a retry hint. A Telegram outage parks the alert in SQLite instead of dropping it.

Runs from a clean clone in one command

Met

Verified by cloning the public repo to a temp directory. That check found a real bug: box-drawing characters crashed the demo on a cp1252 console.

A real eval with committed results

Met

37 cases with results committed as JSON and Markdown. Scored for unsafe decisions, not just accuracy, and two cases are left failing on purpose.

Decisions can be defended in writing

Met

Two ADRs with rejected options and what each choice cost, plus a post-mortem on a bug that only appeared outside my own machine.