Résumé

Paul Jialiang Wu, PhD

Strategic Collaboration Lead & Advanced AI Scientist · Physical AI, World Models & Agentic Systems · Frontier-AI Ecosystems

wjlgatech@gmail.com · github.com/wjlgatech · linkedin.com/in/paul-jialiang-wu-phd · live portfolio ↗

SUMMARY

High-agency, founder-mindset engineer who ships bespoke agentic AI inside customer environments. As Accenture’s Strategic Collaboration Lead and an Advanced AI Scientist & Engineer — and earlier, a Forward Deployed Engineer on an OpenAI enterprise engagement and technical lead on its Physical AI team — I take generative AI from rapid prototype to production-grade reality — architecting the connective tissue between frontier models (Gemini-class VLMs, GPT-class models, NVIDIA Cosmos) and live customer infrastructure, then hardening it with evaluation and observability pipelines for accuracy, safety, latency, and cost-per-request. Creator of open-source multi-agent systems (loop-engineering-anything, super-u) that generate MCP-/agent-native tooling, grade it against reality with an independent referee, and refactor to convergence under hard safety gates. PhD-trained, with 8+ years delivering production ML/AI for Fortune 500 clients; fluent in multi-agent orchestration (ReAct, self-reflection, hierarchical delegation), RAG over structured and unstructured data, and the LLM-native metrics that decide whether an agent is enterprise-ready.

RECENT OPEN SOURCE — JUNE 2026

12 active repositories, 280+ commits in June 2026 — all public at github.com/wjlgatech. Spanning agentic loops, model-quality evaluation, and token/latency optimization.

loop-engineering-anything*Self-Improving Agentic Systems · Creator · 88 commits in June 2026*

• Open-source loop orchestrator (Python, 286 passing tests, CI on 3.11–3.13) that turns any API or codebase into a self-improving, agent-native CLI by driving a generate → judge → refactor → re-judge loop to convergence — walk away, return to a measurably better tool plus an audit report of how it improved.

• Quality is set by an independent referee, not model self-report (a “maker ≠ checker” evaluation pipeline): 5-dimension A–F grading with a hard safety gate that makes unsafe output a terminal, unshippable state regardless of other scores.

• Multi-signal convergence policy — plateau detection \+ regression rollback \+ iteration/token budget — prevents recursive agent degradation and enforces cost limits; every run, grade, and learning is persisted to SQLite for cross-run tracing.

• Fleet layer coordinates many loops under one goal with dependency-ordered execution and automatic feedback routing, escalating only high-judgment forks (safety block, stuck worker, gated result) to a human.

super-u*Multi-Agent Human-Upgrade Platform · Creator · 59 commits in June 2026*

• Multi-agent application (FastAPI, spec-/guardrail-driven CI) across three agentic layers — Creator, Skillfy, Digital Twin — behind a single LLM-provider seam (Anthropic or any OpenAI-compatible host, incl. NVIDIA NIM) for portable model swaps.

• Skillfy is a RAG-style ingestion-to-tool pipeline: ingest text / website / PDF (incl. arXiv) / YouTube transcript / folder → triage gate (quality floor \+ overlap dedup against a knowledge graph) → edge-gated extraction → export as an installable agent tool (SKILL.md / MCP).

• Digital Twin is an observability layer for agent alignment/drift: dual heuristic \+ LLM-judge scoring against a vision kernel, plus per-call efficiency telemetry that meters every LLM call — the heuristic floor any model or prompt change must beat.

More shipped that month — same generate → judge → ship discipline:

• FDE-os — Forward-Deployed-Engineer OS: runnable Claude MCP server, RAG-grounded guide, and an eval-loop \+ criteria-scorer behind a live agentic webapp. *(25 commits)*

• loop — Desktop AI overlay with multi-LLM failover tiers that detect a dead model before the user does, plus one-click local (Ollama) inference. *(17 commits)*

• dreammaketrue — Any source → interactive knowledge graph; engineered 10× perceived AI latency via streaming, caching, and model pre-warm. *(17 commits)*

• cli-judge — LLM-as-judge scoring harness for CLI agents — a reproducible Definition-of-Done scorecard for ranking model/agent output quality. *(15 commits)*

• agentic-portfolio — Conversation-editable portfolio (owner-gated, voice input) with a self-healing, class-independent web harvester. Live app ↗ *(10 commits)*

• Google5Days-Agentic-Engineering — Developer-enablement curriculum with a deployable guide agent (swappable Gemini/Claude backends) and a tested hands-on track. *(7 commits)*

• External OSS: merged a self-healing 429-fallback fix into NousResearch/hermes-agent — honoring api\_key\_env across the fallback chain so agents recover from rate limits instead of failing.

EXPERIENCE

Strategic Collaboration Lead, Adv. AI Scientist & Engineer — AccentureJan 2024–Present · San Francisco Bay Area

*Strategic Collaboration & Ecosystem Building*

• Lead strategic collaboration and branding across frontier AI ecosystems, with direct partnerships spanning research labs and deep-tech startups emerging from Stanford, Berkeley, Yale, Harvard, and MIT.

• Translate these relationships into enterprise innovation, thought leadership, and joint initiatives connecting research, startups, partners, and clients.

• Technical foundation spans Physical AI, NVIDIA Isaac Sim, synthetic data, MCP infrastructure, agentic systems, and Azure ML — bridging frontier technology with enterprise strategy, ecosystem building, and market-facing narratives.

*Forward-Deployed Delivery · OpenAI-DSI Engagement*

• Primary technical interface for an executive-level enterprise AI deployment — recognized by a Senior Manager for upholding a ‘no surprises’ standard and earning go-to status as first point of contact, built on deep model understanding and contextual judgment at decision points.

• Led technical discovery and white-glove deployment: translated ambiguous customer requirements into shipped agentic workflows, and co-built with customer engineering teams to instill rigorous development practices and drive end-user adoption.

*Agentic Systems at Scale · WorkflowX / Agenticom*

• Designed and shipped a multi-agent business-process automation engine for enterprise clients: event-sourced state management for concurrent-agent consistency, idempotent retry across distributed agents, and full audit trail — live autonomous workflows with zero-downtime replanning.

• Solved the core agentic failure mode — agents over partially observable state must replan without losing context — with an event-sourced consistency model that doubles as a granular tracing / observability surface.

*Evaluation & Foundation Models · Multimodal / Physical Domains*

• Built a multi-model evaluation framework comparing 6+ vision-language models (incl. Gemini-class VLMs, NVIDIA VSS, Cosmos-Predict2) on accuracy, latency under real-world load, failure modes on edge cases, and integration cost — the basis for production model selection.

• Shipped a confidence-calibrated selective-verification pipeline — a closed-loop observability system that routes only high-uncertainty outputs to human review, cutting annotation cost \~90% while preserving coverage where model confidence is low.

• Technical lead on 4 simultaneous production systems: synthetic data generation, zero-shot safety detection, adaptive evaluation under distribution shift, and hybrid vision-language reasoning over live sensor streams.

Founder & AI Builder — OmegaFounders / 12XFutureDec 2022–Present · San Francisco Bay Area

• 12XFuture and OmegaFounders are my venture and collaboration platform for one North Star: Making My Ceilings Your Foundations for 12XFuture Together.

• Connect frontier research, deep-tech startups, industry partners, and builders to turn emerging AI capabilities into shared platforms, products, and ecosystems.

• Current focus is AI-native devices: building the data pipelines, agentic infrastructure, and operating systems required to build, train, and deploy World Models.

• With strategic partners from Stanford, Berkeley, Yale, Harvard, and MIT, turn these collaborations into ventures, demonstrations, content, community, and distribution — so breakthroughs do not remain isolated research, but become foundations others can build upon.

Principal Data Scientist — GenentechJun 2021–Dec 2022

• Built and shipped production ML for biomedical research at scale, including 3 open-source PyPI packages for multimodal learning (tf\_multimodal, fastai\_multimodal, fast\_tfrs) adopted by the broader research community for large-scale signal fusion.

• Drove an 85% reduction in unproductive process time via AI workflow optimization; 5-star-rated delivery across all engagements.

Principal Data Scientist — Galvanize Inc.Sep 2019–Jun 2021

• Technical lead across 13+ end-to-end ML systems for Fortune 500 clients — time-series forecasting, NLP, anomaly detection, graph optimization, and large-scale recommendation — owning scoping, stakeholder alignment, technical discovery, and production delivery; 5-star-rated.

SELECTED PROJECT

DataCenterAR — Spatial AI deployed to Fortune 500 field techniciansAccenture Physical AI Team

• Agentic spatial-AI system (Meta Quest XR, real-time monocular depth, affine registration, temporal consensus filtering) that turns physical environments into an agent-navigable coordinate space; architecture evolved from rule-based to fully agentic — clients reported 10x improvement, ‘a coach, not a menu.’ Deployed to Walmart, Costco, and hyperscale data-center clients.

PUBLICATIONS & RESEARCH

SCWM: Self-Calibrating World ModelsNeurIPS 2026 (under review)

• Wu, P.J. et al. — online Bayesian calibration for sim-to-real transfer; 40% reduction in transfer error vs. domain-randomization baseline.

Physical AI: The Next Frontier in AI and RoboticsPreprints.org · Apr 2026

• Wu, P.J. et al. — DOI: 10.20944/preprints202604.0549.v1.

Failure Benchmarking of NVIDIA's VSS Tool: Insights from Vision-Language EvaluationWorking paper · first author · in preparation (no venue selected)

• Wu, P.J., Shah, A., Hosseini, P., Shahab, S. — Center for Advanced AI, Accenture. A five-mode VLM failure taxonomy made executable over deterministic synthetic stimuli, with the Temporal Grounding Score defined as data. Baseline VSS fails 5/5 modes; two frontier models pass 5/5 and the smallest reproduces 3/5 — so the failures track capability, not only late-fusion architecture.

Adapting the BARE Framework for Synthetic Data Generation in Vision-Language ModelsWorking paper · first author · in preparation (no venue selected)

• Wu, P.J., Shah, A., Hosseini, P., Shahab, S. — Center for Advanced AI, Accenture. Base-model diversity \+ instruction-tuned correctness as a twin-gated loop. Mode collapse reproduced (0.23 vs 0.62); base-model hallucination did not. Our own first negative was retracted as underpowered (0/18 ⇒ 95% CI [0, 16.7%]) and re-resolved at n=66 and n=72 as a genuine one.

Eval gates as training-signal factories (RewardForge)Working paper · unpublished, not submitted

• Every honest eval gate both certifies quality and labels preference data. 162 pairs from certified gates, zero human labels → DPO from scratch → LoRA → held-out hallucination 0.398→0.287; five seeds against a pre-committed criterion, treatment \+0.098±0.028 vs shuffled-label control −0.046±0.055.

Evidence-gated benchmarking of agentic tool use (MCP-Arena)Working paper · unpublished, not submitted

• 12 tasks × 6 failure-mode categories over real MCP servers, asserting the tool-call trace rather than the prose, so a fluent answer with no tool calls scores zero by construction; ships a hallucinating mock that must score 0 in CI.

Predicting a frontier lab's post-training stack, scored rather than assertedWorking paper · 12 predictions registered, Brier UNSCORED

• A ten-stage reconstruction of a frontier lab's training stack read from its publications, with 12 falsifiable claims registered 2026-08-13 and resolving by 2026-12-31 against the eventual technical report. 0 of 12 resolved, so the Brier score is reported UNSCORED rather than estimated.

The zero that wasn't evidence — power discipline for gates that report negativesMethods note · unpublished, not submitted

• An audit of our own false negative and the rule it produced: wherever a gate reports a negative it must state the n and the smallest effect that n could have resolved — with a runnable tool (Wilson intervals, rule of three, minimum detectable difference), because a rule without one is a slogan.

WRITING & OPEN SOURCE — REINFORCEMENT LEARNING & SELF-IMPROVEMENT

Reinforcement Learning — a five-part seriesAgentic portfolio · Aug 2026 · each episode with a CPU-runnable lab

• The Consequence Engine — REINFORCE, RLOO, PPO, GRPO, DAPO, Dr.GRPO, GSPO, CISPO and RLVR read as one loop with seven boxes, most of the acronyms arguing about one term: how much credit a behaviour deserves.

• The Probability Forge — policy-gradient baselines and variance: removing the baseline keeps the estimator unbiased but makes it expensive, and a variance problem raises no error.

• The Reward Is the World — reward hacking as Goodhart's law: the question changes from "is my reward correct?" to "how much optimization pressure can it survive?"

• The Scale Furnace — silent failure modes of RL at scale: entropy collapse, groups returning identical rewards so the advantage and the gradient go to zero, and GRPO length normalisation inflating the length of incorrect answers.

• When the Model Gets Hands — agentic RL: the harness becomes part of the policy, an agent rewarded for calling the test runner calls it instead of editing the file, and the repair is rewarding state change.

Loop Engineering — self-improving loops and their refereesAgentic portfolio · Aug–Sep 2026

• Loop Engineering Is an External Referee, Not a Retry — a loop improves only when the scorer sits outside the generator and every accepted step is a measured non-regression; reconciles Reflexion and Self-Refine with the finding that models cannot self-correct reasoning without external feedback; ships a free Colab lab.

• I Scored 100% on My Own Checklist. That Was the Bug. — my own engine scored 23 of 23 on a rubric I wrote after an hour inside its codebase; rewritten by a hostile criterion (include a check only if I expect it to fail), the honest score became 25 of 28 with the gate red.

• rsi-os — open-source living resource on recursive self-improvement and automated AI research: papers, tools, people, labs, benchmarks and learning roadmaps, cross-linked as a knowledge graph with automated weekly freshness checks.

• flywheel — closed-loop self-improvement substrate: flywheel(X) = X self-improved.

TECHNICAL SKILLS

Agentic Systems & Multi-Agent: Multi-agent orchestration (ReAct, self-reflection, hierarchical delegation), MCP servers, event-sourced state management, agentic replanning, idempotent retry, full audit trail; LangGraph / CrewAI / ADK-class patterns

Generative AI & Foundation Models: Gemini-class VLMs, GPT-class models, Qwen-VL, NVIDIA Cosmos / VSS; prompt & context engineering, multi-model evaluation, confidence calibration, distribution-shift detection

Data, Retrieval & RAG: RAG architectures, vector databases, embeddings, structured \+ unstructured data pipelines, document / PDF / web / transcript ingestion, dedup against knowledge graphs

Evaluation & Observability: Independent-referee evaluation pipelines, multi-dimension grading, LLM-native metrics (tokens/sec, cost-per-request, latency), granular tracing, regression rollback, safety gating

Cloud & Infrastructure: Google Cloud Platform (hands-on, ML projects) & Gemini model evaluation; Azure ML (production); NVIDIA AI stack (NIM, Cosmos, VSS); Python, PyTorch, FastAPI, Docker, Kubernetes, CI/CD, SQLite / Postgres, React, WebSocket

Forward-Deployed Delivery: Technical discovery, executive stakeholder alignment, white-glove deployment, prototype → production, customer engineering enablement, field-pattern → reusable module / product feedback loop

EDUCATION

Yale University — National BioMed Fellow, Computational Immunology

Georgia Institute of Technology — PhD, Bioinformatics

University of South Carolina — MS, Mathematics & Computer Science

Sun Yat-Sen University — BS, Applied Mathematics