Selected work, three eras.

Problem

Owners and senior staff were still the “system of record” for admin: reading PDFs, re-keying into spreadsheets and line-of-business tools, chasing missing fields, and drafting the same email variants. Headcount for pure admin was hard to justify; growth work kept slipping because the boring work never stopped. Off-the-shelf RPA demos looked cheap and failed on exceptions; an ungoverned ChatGPT tab created risk with no audit trail.

Approach

• Mapped the real office day: which tasks burned hours, which were rules, which needed judgment — and which the owner insisted on seeing personally.
• Picked 2–3 high-volume, high-pain workflows first (document/email intake → structured fields → queue or human exception), not a company-wide “AI transformation.”
• Chose the right coworker for each surface: Copilot Cowork inside the Microsoft 365 estate, Claude Cowork for long-document and multi-step desk work, Grok where speed and broad reasoning helped — never one model forced everywhere.
• Deterministic workflow backbone (Durable Functions / M365 connectors) with LLMs only for extraction, classification, and draft language — never as the only control plane.
• Human-in-the-loop on money, customer-facing sends, and anything irreversible; full audit of what the model proposed vs what shipped.
• Measured hours saved and error rate on a pilot team before expanding to other locations or process cousins.
• Handoff: runbooks, owners, and cost caps so the automation is an ops asset — not a consultant retainership.

Outcomes

• Pilot workflows cut repetitive admin time dramatically — owners and leads reclaimed calendar for sales, delivery, and expansion work.
• Faster cycle time on intake-to-system: fewer “stuck in inbox” items and cleaner data into systems of record.
• Exception queues replaced silent failure; staff reviewed edge cases instead of re-doing every item by hand.
• Clear ROI story for the board: hours returned per week × loaded cost, minus model and platform spend — the class of AI use case industry reports consistently rank among the strongest returns.

Stack

Problem

Leadership had seen the noise: agents that “run your SaaS,” OpenClaw-style gateways for broad messy tasks, Hermes for fast execution, memory, and scheduled skills. Experiments on laptops worked for a week, then stalled — no tenancy story, no approval path for customer email or CRM writes, no one accountable when the agent loop burned tokens or took a wrong action. They needed help turning agent hype into day-to-day operational leverage.

Approach

• Inventory which day-to-day activities are agent-ready: inbox triage, research briefs, status updates, content drafts, CRM hygiene, recurring reports — vs what must stay human.
• Choose fit, not fashion: OpenClaw when a multi-channel / broad automation gateway and quick path-to-running matter; Hermes when multi-agent skills, persistent memory, scheduled jobs, and tighter execution loops fit better — often both in complementary roles.
• Design tool surfaces with least privilege: which APIs the agent may call, what is read-only, what needs human approval before send or write.
• Wire scheduling (cron / heartbeats) for work that should happen without someone opening a chat window every morning.
• Add observability, cost caps, and kill switches; store secrets in proper vaults — not .env files on a founder laptop.
• Pilot one or two live workflows with success metrics (hours saved, cycle time, error/escalation rate), then expand only what proves out.

Outcomes

• Agents handling a defined set of day-to-day activities on a schedule — with humans reviewing exceptions instead of doing every step.
• Clear split of labor: OpenClaw and/or Hermes where each is strong, not a religious single-platform bet.
• Security and ops baseline: approvals, logs, identity, and an owner who can disable the agent without a war room.
• A path the internal team can maintain — Hyper Mind designs, ships, and hands off; optional fractional support after go-live.

Stack

Problem

Deal closed on a fixed calendar. Two tenants, two identity directories, overlapping SMTP domains, and a business expectation that “after the weekend we are one company” — not a six-month dual-tenant purgatory. Leadership would tolerate a short, planned freeze; they would not tolerate weeks of “check the other tenant” for mail and files.

Approach

• Treated the weekend as a flip, not a discovery window: inventory, identity mapping, UPN/SMTP plan, and coexistence rules locked weeks ahead.
• Pre-staged bulk mailbox and OneDrive/SharePoint content into the target tenant so cutover was delta + DNS/identity, not a cold copy of years of data.
• Domain cutover runbook: remove domain from source, verify at target, MX/Autodiscover/SPF/DKIM/DMARC sequenced with owners on a bridge call.
• Mail flow and free/busy tested in pilot cohorts; Teams and SharePoint path decisions explicit (migrate vs rebuild collaboration surface).
• Comms pack for end users: what changes Monday, what stays the same, where old mail lives, who to call — shipped before Friday close of business.
• Rollback criteria written in advance (even when the preferred path was commit-forward) so the bridge room was not inventing policy at 2 a.m.

Outcomes

• Acquired company users authenticated and worked in the parent Microsoft 365 tenant Monday morning with mail and core collaboration available.
• Historical email and file data available post-cutover — not “we’ll migrate archives later” left as a surprise.
• User-facing disruption limited to a planned weekend window; no multi-week dual-primary tenant for day-to-day work.
• Playbook documented for subsequent M&A moves: same stages, different seat counts.

Stack

Problem

Prototype answered well on curated docs and hallucinated on customer tenants. No eval set, no tracing, no clear ownership of grounding vs generation.

Approach

• Scoped the first shippable path: retrieval-augmented answers with citations, not open-ended multi-agent sprawl.
• Built a golden set from real support tickets (redacted) + LLM-as-judge + human spot checks.
• Agent tools only where deterministic APIs existed; human-in-the-loop for irreversible actions.
• Observability: traces per request, prompt/version tags, cost and latency budgets in the release checklist.
• Feature flag + shadow mode against production traffic before customer-visible rollout.

Outcomes

• Feature shipped behind a flag with eval regression gates in CI.
• Hallucination rate on the golden set dropped to an agreed threshold before GA; citations required for product answers.
• Runbooks and ownership handed to the client’s eng team — no black-box retainership required to stay up.

Stack

Problem

On-prem farms (mail, collaboration, DR) were expensive and aging. A pure lift-and-shift quote looked cheaper on day one and more expensive every month after.

Approach

• Workload triage: rehost / refactor / retire — with explicit “stays on-prem for now” decisions.
• Hub-and-spoke landing zone, private endpoints for data plane, Entra ID as the control plane.
• IaC (Bicep) for repeatable environments; no snowflake subscriptions.
• FinOps from week one: tags, budgets, reserved capacity where utilization was proven, rightsizing loops after 30/60/90 days.

Outcomes

• ~35% average infrastructure cost reduction across the hybrid program (retired farms + rightsizing + license consolidation).
• Documented target architecture the internal team could operate without us in the critical path.
• DR simplified: fewer dual-datacenter ghosts, clearer RTO ownership.

Stack

Problem

Legacy multi-domain Exchange, brittle federation, and a political calendar that forbade a single big-bang cutover. Prior “assessments” had stopped at mailbox counts.

Approach

• Assessment that treated network egress, ADFS cert lifecycle, and UPN cleanup as first-class work — not footnotes.
• Pilot cohorts with real rollback plans; hybrid coexistence until each wave was boring.
• Runbooks for free/busy, autodiscover priority, and on-call ownership during every migration window.
• Stakeholder cadence: weekly risk register, not slide decks of green checkmarks.

Outcomes

• 15,000+ mailboxes moved with zero data-loss incidents attributed to the migration path.
• Hybrid coexistence held for the long tail without becoming permanent dual-run debt.
• Playbook reused on subsequent agency and enterprise programs.

Stack