Systems in production, not slide decks.

The challenge

Northwind's investment operations team was manually triaging 800+ deals a day across three legacy systems. Analyst time was 70% spent on data assembly and only 30% on judgment. Turnaround from deal close to memo was 4 business days, with a 6% error rate that required manual reconciliation.

What we did

1. 01Mapped the deal-memo process end-to-end with stakeholders and identified 14 distinct decision points where automation could remove handoffs.
2. 02Architected a 6-agent system (intake, enrichment, drafting, validation, escalation, audit) with clear ownership boundaries and explicit escalation paths.
3. 03Built offline eval suites from 12 months of historical deal data, then deployed behind a feature flag with shadow runs for 2 weeks before cutover.
4. 04Instrumented every agent step with traces, cost ceilings, and guardrails. Hard limits: $0.50 / deal, 30s p95, 0 PII leaving the perimeter.

The challenge

Helios runs 22 clinics across 3 states, each with its own EHR. Patient referrals required 14 manual steps across 4 systems, took 6 days average, and frequently lost context between handoffs. The compliance team needed a system that could pass external HIPAA audit with zero rework.

What we did

1. 01Built a HIPAA-compliant orchestration layer on Temporal with end-to-end encryption and per-tenant key isolation.
2. 02Modeled the referral as a stateful workflow with explicit retry semantics — partial failures resume from the last known-good state, not from the beginning.
3. 03Implemented edge redaction (PHI never leaves the secure perimeter) with tokenization for any external AI calls.
4. 04Wrote the audit story before writing the code — every state transition logs actor, action, input hash, output hash, and timestamp.

The challenge

Stratus's customer portal was stitched together from 4 separate vendor tools — different auth, different billing, no shared data layer. Support tickets took 4x longer to resolve because reps had to context-switch between systems. Renewal conversations suffered.

What we did

1. 01Chose a single Next.js 16 app with server components, Postgres + row-level security for multi-tenancy, and Stripe for billing — boring tech, on purpose.
2. 02Designed a permission model that maps 1:1 to customer orgs and roles, with audit trails on every privileged action.
3. 03Built a real-time ops dashboard using server-sent events and materialized views — sub-100ms p95 on a single-region Postgres.
4. 04Shipped feature-flagged, dark-launched to 5% of tenants, then ramped weekly with telemetry-driven rollback criteria.

The challenge

Quanta's researchers spent 60% of their time on literature review. Existing off-the-shelf RAG systems hallucinated citations and had unbounded cost — a single deep-dive question could cost $4 and take 90 seconds. They needed accuracy AND a hard cost ceiling, not one or the other.

What we did

1. 01Built a hybrid retrieval pipeline: lexical + dense + re-ranked, with citation extraction as a first-class output.
2. 02Used cascade routing — small models for filtering, large models only for synthesis — to keep cost down without sacrificing answer quality.
3. 03Wrote evals against a 5,000-question golden set curated by Quanta's senior researchers, then shadow-ran for 3 weeks before turning off the legacy system.
4. 04Implemented per-task cost and latency budgets enforced by the orchestrator — every request has a hard ceiling, and overruns trigger automatic fallback to a cheaper path.