Customer Support · Implementation Sprint
Tier-1 support agent for a B2B SaaS platform
Problem: Support team handling thousands of repetitive how-to and account-state tickets per week. Agent fatigue and slow time-to-first-response on simple cases.
What we built: A retrieval-augmented support agent that handles tier-1 issues end-to-end inside the existing helpdesk. Confidence-thresholded escalation to humans, full audit trail, and customer-state grounding (account, plan, recent activity) so answers reflect the actual user.
How it ships safely: Eval suite of 250+ representative tickets. Shadow-mode launch reviewing every response against a human reply for two weeks. Gated rollout per ticket category with rollback runbook.
Retrieval
Eval-driven rollout
Helpdesk integration
Audit trail
~60% tier-1 tickets handled without escalation
< 30s typical first-response time
100% responses logged with retrieval citations
Document Intelligence · Implementation Sprint
Contract review and clause extraction for a procurement team
Problem: Procurement reviewing hundreds of vendor contracts per quarter. Manual clause extraction, redline drafting, and risk flagging consuming senior legal hours.
What we built: A document intelligence pipeline that ingests PDFs and DOCX, extracts a fixed set of clauses (term, renewal, indemnity, liability cap, data handling, AI use), flags policy deviations against the company's standard playbook, and routes flagged items to a human reviewer with a redline draft.
How it ships safely: Per-clause precision/recall evals on a 200-document gold set. Reviewer-in-the-loop UI with confidence badges. Full provenance from extracted clause back to source page.
Doc parsing
Clause extraction
Policy diffing
Reviewer UI
~70% review hours reclaimed on standard contracts
> 95% precision on flagged risk clauses
0 auto-approvals — human always confirms risk
Sales Workflow · Fractional Engagement
Sales call analysis and follow-up drafting
Problem: Sales reps losing context across multi-call deal cycles. Inconsistent CRM hygiene and slow follow-up drafting after discovery calls.
What we built: Post-call analysis that pulls transcripts from the call platform, extracts buyer signals, generates a structured summary into the CRM, and drafts a tailored follow-up email per call. Reps review and send; the model never auto-sends.
How it ships safely: Reps grade outputs in-app for ongoing eval. Quarterly comparison of follow-up quality vs. baseline. CRM writes are reversible and logged.
Transcripts
CRM sync
Draft-only output
In-app grading
~3x follow-up speed vs. manual drafting
~80% drafts sent with light edits only
0 auto-sent emails to customers
Internal Knowledge · Implementation Sprint
Internal knowledge assistant over policies and SOPs
Problem: New hires and existing staff repeatedly asking IT, HR, and operations the same set of policy and process questions. Tribal knowledge stuck in PDFs and wikis.
What we built: Internal Q&A assistant grounded in approved policies, SOPs, and FAQs. Answers always cite the source document and section. RBAC-aware so HR-confidential content is only available to authorized employees. Slack and intranet entry points.
How it ships safely: Pre-launch eval against a 300-question gold set across IT, HR, finance, and ops. Confidence threshold with "I don't know — talk to [team]" fallback. Monthly content freshness audit.
RAG
RBAC
Slack
Source citations
~50% repeat questions deflected from team inboxes
100% answers carry source citations
< 1% "I don't know" responses without follow-up route
Governance · Enterprise AI Governance Engagement
AI governance and security review for a regulated enterprise
Problem: Enterprise launching its first wave of AI features and unsure how to clear security review, compliance, and board-level oversight. Engineering shipping faster than governance can keep up.
What we built: A governance program covering AI policy, vendor and model approval, eval and monitoring requirements per use case, an internal "AI launch checklist," and a quarterly executive briefing format. We staged it as a 4–6 week engagement so engineering didn't lose momentum.
How it ships safely: Policy aligned with existing security and compliance frameworks (SOC 2, HIPAA where applicable). Launch checklist piloted on two real systems before company-wide rollout.
AI policy
Vendor review
Eval gating
Exec briefing
4–6 weeks to a documented governance program
2 systems piloted through the new launch checklist
1 quarterly exec briefing format adopted