Delivered Systems and Bounded Engineering Research

Challenge

In the era of fully-automated AI lock-picks (Mythos and its successors), enterprises can no longer rely on once-a-year scans or one-shot external pentests. BMCs — the most sensitive, lowest-level server management surface — demand a 24/7, self-hosted, self-scheduled red-team loop. Generic agent shells ("OpenClaw + raw LLM") aimed at production BMCs are unguarded bombs on physical hardware.

Execution focus

• 268 specialized red-team skills routed across 7 assessment lanes and 11 vendor / family packs — Unauth-DAST, Web UI deep parsing, Redfish privilege mapping, IPMI / OEM command audit, iKVM, supply-chain SBOM, the OpenBMC / libpldm collaboration track, and a QEMU/OpenBMC emulated-validation lane.
• Multi-model Council architecture: a Hunter (Claude Opus 4.8) and a Skeptic (GPT-5.5) debate every candidate, and a deterministic Governor makes the final Fail-Closed call — neutralizing single-model hallucination and prompt-injection contamination.
• Proof Ladder + evidence priority: every finding is promoted step by step — static candidate → deployed daemon reachability → QEMU full-system / lab_validation_emulated → approval-lease-gated real-hardware validation. No "AI memo for executives" passes as a security conclusion; every high / critical finding ships with a replayable evidence pack.
• Latest platform pass: BMC-TLO multi-step emulation benchmarks, automatic chain construction, fleet-scale raw / unique / reachable / confirmed_emulated counting, throughput SLOs, tool-index preflight, field journals, redaction checks, and knowledge-base poisoning scans.
• ADR 0002 purple closed-loop discipline: every offensive primitive auto-generates a paired Sigma detection rule that lands in the SIEM in parallel with remediation. Dangerous phases require explicit --allow-* approval + watchdog preflight / after-action + a global PANIC halt; confirmed vulnerabilities trigger an automatic coordinated-disclosure timer.
• Field result: the OpenBMC `libpldm decode_get_types_resp()` report progressed from a source-level OOB read to deployed pldmd reachability, controlled fake MCTP peer-path evidence, and a candidate fix shape — the OpenBMC security team responded and indicated they will address further issues in the area. The same completion-code-first pattern extends to FRU / BIOS / Platform / Firmware Update / IBM OEM decoder hardening candidates.

Impact

Challenge

Natural-language bare-metal operations need vendor compatibility without trading convenience for uncontrolled write access.

Execution focus

• Implemented vendor profiles and an abstraction API surface for Redfish endpoints, supporting local development and validation of cross-generation operational workflows.
• Separated capabilities into read-only, safe-write, and full-operation control modes, with retries and evidence capture for reproducible operation records.
• Applied per-host risk ceilings, confirmation tokens, rate limits, and audit events to keep write operations within explicit risk boundaries.

Impact

Challenge

Repeated engineering-agent jobs need batch execution while preserving traceable sessions, tool calls, and operational boundaries.

Execution focus

• Implemented the `cx` command-line interface and Pi provider bridge to launch and manage Cursor agent runs.
• Provides persistent sessions, project preambles, and run isolation so each engineering task retains inspectable context and boundaries.
• Captures tool-call and JSONL audit trails, with operator-configured sandbox options for controlled execution.

Impact

Challenge

Translating conceptual hardware descriptions into schematic design files requires repetitive specialist work and verification at every reviewable step.

Execution focus

• Researches parsing, semantic extraction, and visualization for structured and HDL inputs, marking parser certainty with byte-addressable evidence and deterministic gates.
• Built local demo prototypes to explore reviewable analysis and conversion paths between Concept HDL and OrCAD Capture representations.
• Defines validation stages and packaging for reviewable engineering output; the CPM-to-DSN writer remains unproven.

Impact

Environment and constraint

A bilingual live-interpretation workflow for operator-controlled use, bounded to a local workflow, explicit failure handling, and a reviewable acceptance process.

Reusable platform core

Reusable workflow orchestration, deterministic fixtures, archival, and evidence collection provide a consistent foundation for later validation.

Customer-specific module

The bilingual interpretation flow combines operator controls and exception-handling steps without encoding any organization, person, venue, or access configuration.

Acceptance and control boundary

Deterministic fixtures, failure handling, and local archival form evidence-backed acceptance gates; this makes no claim of event deployment, event readiness, accuracy, or latency.

Reusable improvement returned to the platform

Reusable controls, fixtures, acceptance gates, and archival patterns return to the platform without sharing workflow content.

Impact

Environment and constraint

A prototype for event-stream and stateful operational workflows with multilingual intent and response handling; its data are synthetic and its real-operation decision history is empty.

Reusable platform core

Reusable event-stream handling, state management, human decision logging, replay-based evaluation, and guardrails.

Customer-specific module

Intent, response, and human-decision steps can be configured for a specific operational context; this example exposes no customer workflow, integration, or data.

Acceptance and control boundary

Guardrails and replay-based evaluation review prototype behavior; production integrations remain roadmap work, not a claim of real operations, customer deployment, or adoption.

Reusable improvement returned to the platform

Prototype-validated decision logging, replay, and guardrail patterns return as reusable operational-workflow capabilities.

Impact

Environment and constraint

Implemented integration and optimization for GPU-efficient retrieval and reranking services, excluding internal infrastructure, endpoints, hosts, and topology.

Reusable platform core

Reusable embedding and reranking service integration, tests, and operational optimization for retrieval workflows.

Customer-specific module

Embedding, reranking, and operational modules can combine for a particular retrieval workflow without publishing any customer, employer, dataset, or access detail.

Acceptance and control boundary

Described as implemented integration and optimization around upstream models; it neither presents the models as original nor makes production, performance, or compliance claims.

Reusable improvement returned to the platform

Tested service-integration and operational-optimization patterns return as reusable platform capabilities for later retrieval workflows.

Impact

Challenge

The customer needed a controlled AI assistant over web content, SharePoint, PDFs, Office files, and multi-tenant configuration without sending private data into a generic chat tool.

Execution focus

• Built a multi-source ingestion and document parsing pipeline into a unified searchable data layer.
• Supported multi-tenant configuration and zero-code deployment patterns across industries.
• Preserved private deployment, permission control, tracing, and evaluation for long-running production operations.

Impact

Challenge

Keyword search could not answer complex enterprise knowledge questions, while vector-only retrieval created recall gaps and weak evidence.

Execution focus

• Combined semantic retrieval, keyword matching, and RRF fusion over 1M+ vectors and 778K+ documents.
• Added multi-stage reranking, metadata filters, parent/sibling expansion, and multi-hop reasoning.
• Used agents to coordinate database queries, web search, and other tools so static documents and live data could share one answer path.

Impact

Challenge

The early RAG pipeline had high latency and insufficient import throughput, blocking realistic enterprise production usage.

Execution focus

• Optimized retrieval, reranking, context expansion, chunking, and resource scheduling.
• Removed unnecessary calls and improved data import concurrency and batching.
• Established repeatable evaluations so accuracy, latency, and throughput could be tracked over time.
• Folded in GPU-efficient embedding and reranking service operations, tests, and optimization around upstream models; the service and operational optimization are original integration work, while model ownership remains upstream.

Impact

Challenge

The team had large volumes of repetitive expert work that needed automation speed while retaining human review and auditability.

Execution focus

• Code-Fix Agent identified, repaired, and verified code quality issues at scale.
• AI Recruiter extracted screening criteria from job descriptions and generated candidate shortlists.
• Kept automated actions and human confirmation in the same workflow to avoid uncontrolled black-box execution.

Impact