Machine Learning Engineer | LLM Inference | Post-Training & RL Environments | Evals & Safety Research

Featured: Latent Space AI in Action Talk, Oct 2025 Technical Blog: The Third State in AI alignment

I build and harden AI systems where failure is expensive. I'm Co-Founder and Head of Engineering at Attunica (clinical AI for psychotherapy training, live on AWS in 2 pilots) and was the first hire at GhostEye (YC S25). On my own I run Chimera, a public inference-safety research program: 55 technical reports / 1.46M+ measurements (47 public), 2 workshop papers accepted (ICML 2026, NeurIPS 2026), seven merged fixes in vLLM, PyTorch, Ollama, and Triton, two PyPI tools with 57K+ downloads, and 4 public RL/evaluation projects with interactive demos.

chimeraforge.vercel.app/projects · 11 systems I built, each rebuilt in the browser on synthetic data so you can run it yourself. Source repos are public.

Also live:

• Agents & evaluation: spreadsheet reasoning inspector · browser-agent completion gate · intake triage scorer
• Systems: staged document search · send pacing scheduler · drone mission governance
• Product: collaborative whiteboard

icml2026-paper-reproductions · Final: rank #26 of 1,221 participants (top 2.1%) · 48 papers · 249 official claims judged: 118 verified, 3 published claims falsified · 297 pts · verdicts frozen 2026-08-02.

Independent, from-scratch reproductions of ICML 2026 submissions, produced by an agentic pipeline and judged claim by claim by the challenge's independent referee. Acceptance and falsification rules are pre-registered before each run; producers cannot publish their own bundles, and reviewers cannot repair what they judge. Every reproduction ships its per-paper evidence: methodology · frozen leaderboard.

Founder & Lead ML Architect · Sep 2025 – Present · New York, USA

Also: Chimeradroid (Unity/C# JARVIS client for Android and Android XR), Echo (Slack, Discord, Telegram, WhatsApp, and email relays), ProjectWyvern (constitutional drone-autonomy layer over PX4/ArduPilot, in simulation), and Banterblogs (write-ups).

Banterpacks, Banterhearts, and Muse Protocol are private during the publication window; read access on request via LinkedIn.

55 technical reports (47 public: TR 108–149, 152, 163–165, 167) · 1.46M+ decision-grade measurements, curated from ~10⁹ profiler samples · 3 of my own pre-registered hypotheses overturned. Every public TR is a markdown file in /reports/: count, read, diff. Pre-registered paired designs, bootstrap CIs, TOST equivalence, Cohen's d, Holm-Bonferroni.

• Safety tax of inference optimization: normalized over two common anchor models, quantization accounts for 57% of the safety-score cost, backend 41%, concurrency 2% (TOST null). Across 18 models / 10+ families, alignment type showed no detectable association (p=0.942); output instability predicted fragility best (r=0.91), and chat-template divergence sometimes exceeded precision effects.
• Quantization safety (RTSI preprint): safety can degrade 13.9× faster than quality under quantization. Offline routing across 45 configurations recovered 76% of the refusal gap with 20% routed to direct safety testing (leave-one-cell-out AUC 0.84); the preprint routes 10/10 hidden-danger configs (Wilson 95% lower bound 0.72). Shipped as the QuantSafe Certifier.
• Post-training: objectives matched to evidence shape (paired debate preferences → DPO/ORPO, unpaired verdicts → KTO, verifiable rewards → Dr.GRPO/RLOO/REINFORCE++) behind shared eval and promotion gates. A pre-registered Dr.GRPO RLVR run (Qwen2.5-1.5B, MedMCQA) traced a null to 76% zero-reward-variance rollout groups; changing only the prompt gained +8.8pp held-out pass@1 (p=0.0003), with late rollout collapse reported alongside.
• Overturned: M/D/1 queueing missed continuous-batching latency by 20.4×; NUM_PARALLEL had no detectable effect (0/30 significant); PyTorch Direct throughput degraded more than Ollama. A four-stack study isolated PyTorch Direct's N=2 dispatch breakdown, removed by TGI on the same GPU; a separate vLLM benchmark reached 2.25× Ollama's throughput at N=8.

Oct 2025 – Present · New York, USA

Clinical AI platform for psychotherapy training: social-work students run sessions with voice-and-avatar AI clients, and instructors assess them. Live in 2 pilots: the NYU Silver MSW program and a 120-therapist clinic; HIPAA BAAs executed across Anthropic and AWS. I architected and solo-built the platform core and lead a PM and two engineers.

^{Walkthrough recorded on a local stack with synthetic practice data.}

• First-customer release live on AWS (Sep 2026): backend, frontend, LiveKit agent, and evaluation services on ECS, with Aurora PostgreSQL 18 and Bedrock; the deployed evaluator was accepted only after an audited Bedrock canary
• Real-time sessions on LiveKit + Gemini Live + Anam avatars, with Deepgram producing the canonical transcript and recorded sessions under revocable consent
• Five-criterion formative evaluation that separates "no evidence" from "scored zero", plus an instructor human-assessment lifecycle (immutable submit, receipted release) with the rubric blocked from automation
• Judge-validity benchmark (40 fixtures, held-out scenarios, bias probes) for three-sample median judging: debiasing moved leniency from +0.81 to within 0.13 anchors of zero (held-out weighted kappa 0.66 vs 0.05 null; model-written key), and showed that sample disagreement alone misses large errors
• Instructor authoring with PDF/DOCX source ingestion: hash-bound uploads, macro and external-link rejection, source text treated as untrusted input
• Every change passes an exact-base validator with append-only admission, so a PR cannot weaken the checks that approve it
• Article 31 documentation product for clinicians (v0.5.1 on ECS): release-gated deploys, on-device Whisper dictation, browser-only PDF extraction, psychotherapy-note authorship enforced end to end

Dec 2025 – Mar 2026 · New York, USA

Built a security awareness training platform in 90 days as a founding engineer. Multi-channel delivery across web, Slack, Teams, SMS/RCS, WhatsApp, Telegram, voice, and email.

• Shipped to 8 pilots (5 enterprise, including a top-10 global asset manager and a Fortune-100 cloud platform); barge-in, streaming, caching, and summarization cut per-turn latency from 5–7s to 0.5–1.5s, and shared caching and summarization libraries cut LLM costs 30–80% across all agents
• Phishing email generation pipeline on self-hosted 70B LLMs with domain-specific LoRA/QLoRA adapters trained with DeepSpeed on a 1M+ email corpus
• Reduced deepfake phishing simulation latency from 40s to 100–450ms (80–400x improvement) via a multi-agent WebRTC pipeline (video render agent + voice agent); range reflects per-call workload depth

Languages: Python, TypeScript, Rust, C#, SQL, C++

ML & inference: PyTorch, Transformers, DeepSpeed, vLLM, SGLang, TGI, TensorRT-LLM, llama.cpp, CUDA, Triton, FlashAttention, torch.compile, Nsight

Post-training, RL & evals: LoRA/QLoRA, DPO/ORPO/KTO, GRPO-family RL, REINFORCE, RLAIF, Gymnasium environments, offline policy evaluation (FQE, doubly robust), PRM/ORM routing, TOST, bootstrap CIs

Product & infra: FastAPI, Next.js, React, PostgreSQL, Redis, ClickHouse, AWS (ECS, Bedrock, Aurora), Terraform, Docker, Kubernetes, LiveKit, OpenTelemetry

A Paired Testing Protocol for Batch-Conditioned Refusal Robustness in LLM Serving Accepted, ICML 2026 Workshop on Hypothesis Testing

A Safe Prototype Is Not a Safety Direction: Reference Dependence and Prompt Confounds in Response-Safety Embeddings Accepted, NeurIPS 2026 Workshop on Foundations of Language Model Security (FLMSec)

Quality Is Not a Safety Proxy Under Quantization: The Refusal Template Stability Index Preprint

Speculative Decoding at Temperature Zero: A Scoped Safety-Invariance Screen with a 48,072-Sample Expansion Preprint

NeurIPS 2026 workshops: RTCA program committee (5 reviews), JUDGe (3), FLMSec (2). Ethics reviewer for the NeurIPS 2026 main track and the Evaluations & Datasets track.

Led a 3-engineer, 1-physician team across a 5-institution clinical imaging program: TensorFlow/Keras pipelines over dental, retinal, and EEG data with SHAP interpretability. Registered work: Copyright L-122721/2023.