Forward Deployed Engineer — Applied AI & Agent Systems
I turn ambiguous operating problems into deployed AI systems, then turn the repeated failure modes and integration patterns into reusable infrastructure.
My work spans agent runtimes, MCP/tool execution, voice systems, workflow automation, local inference, verification, runtime reliability, and the business systems around them. The operating pattern is consistent:
customer problem
-> discovery
-> technical scope
-> integration
-> production
-> evaluation
-> hardening
-> reusable platform capability
The deployments are the story. The repos are the evidence.
| Proof surface | What it demonstrates |
|---|---|
| ai-portfolio | Forward-deployed case studies, field notes, resume, proof map, and maturity boundaries |
| Public Runtime Wind Tunnel | Runnable native-vs-governed runtime evaluation with deterministic faults, independent verdicts, evidence bundles, replay, and tamper detection |
| governed-mcp-spine | Pre-execution capability authorization, write-scope enforcement, and tamper-evident audit evidence |
| clue-runtime | Runtime reliability patterns extracted from production incidents: dependency-honest readiness and startup drift detection |
| voxmaestro | Deterministic voice-agent orchestration: YAML state machines, tool bridges, filler gates, and human handoff |
| agentic-crm-os | Multi-tenant lifecycle state and explicit PostgreSQL transition semantics |
For the fastest evaluation path, start with ai-portfolio and run the public Wind Tunnel specimen.
I am most useful where a requirement is still partly operational and partly technical:
- find the actual workflow constraint before choosing the model or framework;
- ship the smallest complete vertical slice rather than a disconnected demo;
- integrate with the customer's existing systems of record;
- separate model output from authority, external effects, and acceptance;
- make failure observable and reproducible;
- convert incidents into regression tests, runtime invariants, or explicit controls;
- extract platform primitives only after the same problem recurs in the field.
Working doctrine: logs are telemetry; evidence proves behavior.
- A health endpoint reported
okwhile its database dependency was unavailable for six days → dependency-aware health/readiness in clue-runtime. - A hash-chained audit trail silently restarted across rotation → rotation-safe chain continuation and regression coverage in governed-mcp-spine.
- Runtime/governance claims were difficult to compare when the harness and verifier moved with the system under test → the Public Runtime Wind Tunnel.
These are useful proof/supporting surfaces, but they are not the headline story:
- agentreceipts — local signed observation records; useful as evidence inputs, not authorization or proof of correctness.
- mcp-audit-plugin — MCP configuration-audit prototype and historical diagnostic work.
- smart-ai-router — explicitly incomplete cost-aware routing scaffold.
I distinguish:
- implemented — code exists;
- tested — defined behavior has automated checks;
- verified — evidence supports the expected property;
- accepted — the designated verifier or gate has authorized the result;
- deployed — the system is operating in its intended environment.
Those words are not interchangeable.
Public repositories intentionally exclude customer data, credentials, private infrastructure, and proprietary policy data. Where a claim has runnable proof, I link it. Where it does not, I label the boundary.
Forward-deployed AI systems and governed agent execution: taking real customer workloads from integration through production hardening, then feeding the resulting failure modes into stronger runtime and verification primitives.
- GitHub: github.com/gabeacosta
- LinkedIn: linkedin.com/in/juan-acosta-a47b7a3a0
- Site: gentic.pro
- Email: gabriel@gentic.pro