I build and ship production LLM applications—multi-agent systems, RAG pipelines, evaluation frameworks, and real-time inference—running on local GPU infrastructure with zero cloud dependency. I own quality end-to-end from prompt architecture and context design through trust boundaries and reliability checks.
I design evaluations against real production scenarios, not synthetic benchmarks, and have enterprise systems architecture experience. My work spans multi-agent coordination, semantic search, multimodal pipelines (audio/text/image/video/music), workflow-driven AI assessment with human-in-the-loop gates, and local inference optimization under real GPU contention.
Experience Level
Work Experience
Education
Qualifications
Industry Experience
- Real-time, event-driven — replies within human conversational latency
- Sandboxed by construction — no write path exists, so it is safe to leave running unattended
- Cites its source; escalates rather than fabricating — no confident-wrong answers
- Runs on local inference — no cloud, no data egress
- Stack: TypeScript, tool-calling LLM, event-driven runtime
An event-driven AI assistant that fields real user questions unattended, in real time — and is read-only by construction: it physically cannot write, commit, deploy, or change data, because those tools simply don’t exist in its toolset. A destructive mistake is impossible, not merely forbidden. It answers only what it has been verified it can answer, cites the source behind each answer, and escalates to a human rather than guessing when something is outside its remit.
An autonomous agent you can actually leave running — because the worst it can do is ask a human.
- Observation extraction: local LLM agents pull structured observations from raw call and SMS transcripts
- Multi-layered memory (short / medium / long-term) — continuity across interactions, the way a human analyst carries context
- Ask in plain English: a natural-language query pipeline with progressive refinement over vector embeddings — no query language required
- All local inference (Ollama) — confidential customer conversations never leave controlled infrastructure; no cloud, no per-token cost
- Full-stack product: Node.js backend, Angular front end, RESTful APIs, RBAC + JWT, admin and user management
- Extensible beyond calls to any business process that benefits from persistent, contextual understanding
A stateful, agentic system that turns unstructured business interactions — phone calls and SMS — into structured, queryable process intelligence. Unlike tools that treat each interaction in isolation, it preserves semantic context and memory across the whole process, so it understands the story, not just the sentence.
Business-process intelligence that remembers — semantic context, not just transcription.
- Front end: Next.js (App Router), role-aware single app, shadcn/ui + Tailwind
- Backend: four TypeScript services (gateway, core, integrations, vector) over tRPC — end-to-end type safety from the SQL query to the React component
- Data: Postgres + pgvector for semantic search; a fully typed query layer (Kysely + Zod); an audit event on every state change
- AI, all local: classification, structured extraction, document OCR, agentic tool-calling, and semantic search — zero cloud, no per-token cost, no data egress
- Audit-grade output: generates regulatory reports validated against the authoritative spec, never fabricating a value it can’t ground
- End-to-end tested (Vitest + Playwright); pnpm monorepo
A production platform built end to end for a regulated, compliance-heavy domain — from the database to the UI, with AI woven through, and every model running on local inference so sensitive data never leaves controlled infrastructure.
One engineer, the whole stack — a real product where the AI is load-bearing, not bolted on, and the sensitive data never leaves the building.
- Ingests uploaded documents and structured inputs
- Assesses against a codified source-of-truth (legislation, policy, rubric)
- Cites the rule behind each decision; declines when evidence is thin — no hallucinated approvals
- Full-stack web app; runs on-prem or in the cloud
An AI system that evaluates documents and submissions against an authoritative ruleset and returns a structured, defensible determination — with a citation to the specific rule behind every judgement, and a hard refusal to assert anything it can’t ground in the source. Built for high-stakes domains — regulatory, tax, compliance, screening — where a confident wrong answer is a liability.
Assessments people can trust and audit, because the system shows its work and knows what it doesn’t know.
- Live speech-to-speech over a real phone line (SIP/PBX)
- Understands intent, handles Q&A, and completes bookings autonomously
- Zero cloud dependency — self-hosted models; data never leaves the box
- Stack: Python, telephony/SIP, local LLM inference, real-time audio
A production voice AI that answers phone calls in real time — natural conversation, answers service questions, and books appointments end to end, with no human handoff. Built on self-hosted inference with zero third-party AI APIs, so it runs at near-zero cost per call and keeps every second of call data on controlled hardware.
A callable receptionist that never sleeps — at a fraction of a hosted solution’s cost.
Hire Mark Lee today
To get started post up your job and then invite Mark Lee to your job.