Problem: Operations team running repetitive multi-step workflows across CRM, helpdesk, and back-office tools — losing 200+ hours/month to manual coordination.

Solution: Designed a LangGraph-based multi-agent system with typed tools, retries, and human-in-the-loop approvals on irreversible actions. Full Langfuse tracing on every run.

Outcome: Automated workflow coverage moved from 0 to 70% of in-scope tasks within a quarter, with a measurable drop in handle time and zero hallucinated actions in production.

Problem: Knowledge sprawl across 100k+ documents — contracts, SOPs, manuals — with brittle keyword search and growing tickets to legal & ops.

Solution: Hybrid retrieval (BM25 + dense + re-ranker), structured extraction, citation-grounded answers, and a regression eval suite on every change.

Outcome: Self-service answer rate measured against a 500-question eval set, with citation grounding above target. Significant reduction in escalations to internal subject-matter experts.

Problem: Recruiters losing hours on first-round screening calls — repetitive questions, repetitive note-taking, missed follow-ups.

Solution: Outbound voice agent for first-round screening with real-time ASR/TTS, structured note extraction, and direct CRM sync. Human handover on edge cases.

Outcome: Recruiter time on first-round screening dropped sharply; structured candidate data flows directly into the ATS with searchable transcripts attached.

Problem: Client meetings producing scattered notes — MOMs late by days, action items missed, decisions disputed.

Solution: End-to-end meeting platform: high-accuracy transcription, speaker diarization, structured MOM extraction, reviewer-approved publishing, and CRM/email sync.

Outcome: MOMs published same day instead of next week. Decisions and actions traceable per meeting with reviewer accountability.

Problem: Legacy IVR with low containment rate and high frustration; live agents repeating the same 10 resolutions all day.

Solution: Real-time conversational layer with strict confidence gating — answers from a curated KB, transfers to a human when confidence drops below threshold.

Outcome: Containment rate up significantly on in-scope intents; live agents freed to handle complex cases. Full call analytics dashboard for ops leadership.

Problem: Manual attendance verification across multiple sites — slow, error-prone, hard to audit.

Solution: On-device face recognition with anti-spoofing checks, encrypted templates, and centralised audit log. Edge-first design for offline resilience.

Outcome: Sub-second on-device verification, fully auditable attendance records, and no plain-text biometric data leaving the device.