“Work that typically took days came back in minutes. Any team building AI for real users would be extremely lucky to have him!”
James NiuSr. Staff AI EngineerOpen to AI engineering opportunitiesSan Francisco · Remote USEMAILLINKEDIN
I build agentic AI systems, then make them hold up under real traffic.
6 months became 5 days. At Basis Technologies, I rebuilt media planning from ideation through the right audience, right time, and right channels, reaching 98% accuracy across 200 turns.
6 months to 5 daysEND-TO-END MEDIA PLANNING WORKFLOW
98%ACCURACY ACROSS 200 TURNS
1.7M tokensCONTEXT COMPACTION, 8X LONGER SESSIONS
I do the whole loop: the agent, the tools it calls, and the evals that decide whether it ships. If you are putting AI in front of real users, that is the part I own.
Media planning workflow cut from 6 months to 5 days
98% accuracy across 200 turns
$175M+ media spend, 3B+ impressions
40 stakeholder recommendations on LinkedIn
AI copilot serving 120+ enterprise advertisers
10 years building production AI systems
01 / STAKEHOLDER PROOF
Trusted across the customer, product, engineering, and quality table.
“He focuses on what happens when the agent is wrong, who is accountable, and how you would know before your customer does.”
“When an agent workflow behaved unexpectedly, James was the person I trusted to diagnose it. Any team building production AI agents would be stronger with him on it.”
“He understood edge cases, welcomed QA feedback, and jumped on bugs early. That commitment to quality showed up directly in our releases.”
“An extremely forward thinker as AI capabilities emerged. A growth mindset with the ability to learn fast.”
ABOUT
Measure the model, not the demo.
Ten years building production systems, the last several on AI agents. I started in quant finance and operations research, so I treat model behavior as a system to measure, not a demo to admire.
At Basis I built Compass, an AI media-strategy assistant over Claude, GPT, and Gemini with LangGraph and MCP, and I owned the eval harness that gated its releases: 6 months of media planning down to 5 days, 98% accuracy across 200 turns.
I work the full stack: Python and TypeScript services, agent orchestration, MCP and A2A tools, and the CI gates that keep agents honest in production.
Based in San Francisco. Open to full-time AI engineering roles, here or remote US.
02 / WORK EXPERIENCE
A decade of measurable AI and business impact.
Basis Technologies
Sr. Staff AI Engineer
Built Compass, an AI media-strategy assistant over Claude, GPT, and Gemini via LangGraph and MCP, and the eval harness that gated its releases.
- 6 months to 5 days for end-to-end media planning.
- 98% accuracy across 200 turns.
- 0.03 false positives across 190+ evals.
Advantage Solutions
Principal AI Solutions Architect
AI systems for retail analytics and workforce operations.
- 40% staffing accuracy.
- 20% campaign ROI.
- 70% lower reporting cost and runtime.
Leadership Circle
AI Solutions Engineer
Customer-facing AI support systems.
- 50% fewer support tickets.
- 99% reliability.
- 30% lower infrastructure cost.
Medlytix LLC
Data and AI Architect
Data and AI for healthcare claims and revenue cycle.
- 25% higher recovery efficiency.
- 25% fewer denials.
- 20% better early detection.
GeoAlliance Consultants Pte Ltd
Forward Deployed Engineer (FDE)
Deployed ML on construction projects: contract review and worksite operations.
- 60% faster contract review.
- 15+ late projects prevented.
- 15% higher worksite efficiency.
M Capital Group (MCG)
Machine Learning (ML) Engineer
ML for portfolio forecasting and risk.
- 30% higher forecast accuracy.
- 20% lower risk.
- 50% less model-ops friction.
EDUCATION
2017 TO 2019Columbia University in the City of New YorkMS in Operations Research
2014 TO 2017London School of Economics and Political ScienceBS in Economics, First-Class Honors
ORCHESTRATION · EVALUATION · RELEASESeven systems, with evals and failure paths.
04 / SELECTED WORK
Production AI agents, evaluation systems, and inference services.
FEATURED CASE STUDY
Autonomous Ads Pipeline
Four-engine ad creation gated before paid regeneration, so a marketing team can trust AI with real ad spend.48HUMAN-GRADED EXEMPLARS
28VERSIONS ACROSS 4 ENGINES
10DERIVED GATING THRESHOLDS
7 STEPS · 4 GUARDS · 10 PROBESGitHub
01 / VOICE AI
Realtime Voice Agent
Turn-taking, interruption recovery, and failover for realtime voice. A voice agent that never talks over your customer.
02 / REINFORCEMENT LEARNING
Multi-Agent Drone Navigation
Coordination under motion, drift, and conflict. RL policies that survive contact with other agents.
03 / LLM EVALUATION
Self-Hosted Evals Lab
Prompt strategies measured with confidence intervals. Stop arguing about prompts; measure them.
04 / FINE-TUNING
Turn Detection Model
The model waits when the speaker is not done. 33.1 ms p95, because in voice, latency is a feature.
05 / ENTERPRISE MCP
Atlassian MCP Server
Enterprise tools with identity and audit built in. MCP the way a security team will actually approve.
06 / COMPUTER VISION
Form Recognition Harness
Uncertain form reads route to human review. Vision that knows when to ask a human.05 / HOW I WORK
The checks between a demo and production traffic.
Trace every tool call.
Fail the build on bad evals.
Route around slow or failing models.
Stop risky actions at the gate.
06 / TECHNICAL SKILLS
What actually ran behind these systems.
AGENTS
Agentic LLM SystemsLangGraphModel Context Protocol (MCP)Agent-to-Agent (A2A)Retrieval-Augmented Generation (RAG)Function CallingTool CallingMulti-Model OrchestrationPrompt Engineering
EVALS
LLM EvaluationLLM-as-JudgeGolden DatasetsCI-Gated RegressionGuardrailsObservabilityTracing
APPS
PythonTypeScriptFastAPIReactSQLAlchemyWebSocketspandas
DATA
PostgreSQLRedisSQLMultimodal AIComputer VisionSnowflakeDatabricksVector Databases
QUANT
Operations ResearchOptimizationQuantitative EconomicsEconometricsQuantitative FinanceStatisticsStochastic ModelingSimulationDecision Science
INFRA
DockerKubernetesAWSCI/CDLangfuseMLOpsLLMOpsMicrosoft Azure
MODELS
ClaudeGPTGeminiPyTorchHugging FaceONNXTensorFlowTransformersFine-TuningElevenLabs
07 / NEXT
Need an AI agent that survives production traffic?
Full-time, individual contributor: Staff or Principal AI engineer, or forward-deployed AI work where agents meet real users. San Francisco Bay Area or remote US.