Production systems. Production numbers.

The problem

Medical coding is slow, expensive, and unforgiving. A trained coder spends 15 to 20 minutes per record assigning ICD-10, ICD-10-PCS, and CPT codes, and every error risks denied claims and compliance exposure.

What we built

An AI coding pipeline built on multi-agent orchestration: separate extraction, coding, and validation agents, RAG with hybrid search and LLM re-ranking over the code sets, year-wise CMS code management, and HIPAA-compliant audit trails on every decision.

Stack

The problem

An urgent-care clinic's front desk can't answer every call. Patients calling about appointments, walk-in waits, hours, and locations hit hold queues during the day and voicemail after hours, and the legacy system took 4 to 6 seconds to respond when it answered at all.

What we built

A production voice agent our senior architects delivered at Piedmont Urgent Care, answering patient calls around the clock: appointments, walk-ins, hours, locations. LiveKit Agents with Twilio SIP, Deepgram Nova-2 speech-to-text, Azure OpenAI GPT-4o-mini with RAG over clinic SOPs, ElevenLabs voice, Silero VAD with ML turn detection. HIPAA-eligible infrastructure with end-to-end encryption.

Stack

The problem

First-notice-of-loss interviews are a bottleneck: adjusters conduct recorded statements one at a time, transcribe them, and hand-build claim files, while fraud signals surface weeks later, if ever.

What we built

A voice agent that conducts recorded-statement interviews end to end: dynamic follow-up questions, live entity extraction into a structured claim file, in-call fraud, litigation, and sentiment flags, and post-call SIU-style fraud triage with human review on medium-plus risk. Handles inbound and outbound with answering-machine detection. Every call produces five artifacts: claim JSON, WAV recording, PDF report, quality critique, and fraud assessment.

Stack

The problem

Serving Spanish-speaking customers usually means hiring bilingual staff or losing the calls. English-only agents can't hold a natural conversation across the language gap.

What we built

A real-time Spanish-English bridge: the customer speaks Spanish or Spanglish, the agent reads a live English translation and replies by voice, and the customer hears natural Spanish. LiveKit orchestration with a React dashboard over SSE, and every call captured as JSON plus WAV.

Stack

The problem

Enterprise support queues depend on humans reading every ticket to route it, and misroutes disappear silently into the wrong team's backlog.

What we built

An MCP-architecture triage engine: a LangGraph preprocess-classify-parse pipeline with confidence-based routing that auto-routes at 0.85+ confidence and hands everything below to humans. Takes screenshot and vision input, runs provider-agnostic across OpenAI, Claude, and Gemini, and keeps categories fully config-driven with no redeploys.

Stack

The problem

Security teams drown in alerts. Every investigation means manually writing queries, pulling context, and deciding severity while the clock runs.

What we built

AI-driven alert triage and investigation: ML threat classification, automated KQL query execution, and real-time monitoring and response workflows.

Stack

The problem

Incident triage inside ServiceNow was a full-time manual job: humans reading tickets and guessing the right assignment group, at enterprise volume.

What we built

ML classification of incidents with automatic categorization and routing inside ServiceNow.

Stack

The problem

SWIFT financial messages arrive as dense, structured text that operations teams parsed by hand: slow, expensive, and error-prone at bank scale.

What we built

NLP and NER extraction over SWIFT messages using CRF models: entities pulled automatically, messages classified into 12 categories, 23 entity types extracted, processed in real time.

Stack