Machine Learning Engineer | Constitutional AI | Inference Systems | Empirical Safety Research

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

Methodology built in the public Chimera adaptive constitutional engine — 15+ services, Rust alignment runtime, 54 technical reports / ~1.3M decision-grade measurements (curated from ~10⁹ profiler samples), 3 overturned hypotheses, 1 ICML 2026 workshop paper accepted + 5 submissions under review at a top ML venue, and 4 merged upstream PRs (vLLM #45207, PyTorch #175562, Ollama #16669, Triton #10819) — and applied at production scale at GhostEye (security training platform shipped in 90 days; 80–400x deepfake latency reduction), Attunica (HIPAA-aligned clinical AI; BAAs across Anthropic + AWS), 16 Hugging Face models, Chimeraforge (PyPI capacity-planning CLI, v0.6.0, 483 tests, 4K+ downloads), and quantfit (PyPI safety-tax CLI: quantize an LLM and check it still refuses what it should, v0.1.0). Everything measured, everything reproducible — count the markdown files yourself.

Full-stack: CUDA kernels and Triton compilation up through multi-agent runtimes, alignment architectures, and production platforms. Three pillars:

1. Inference optimization — vLLM, TGI, Ollama, TensorRT, torch.compile, FlashAttention, quantization sweeps, Nsight Systems kernel profiling. I don't guess where the bottleneck is. I trace it.

2. Constitutional AI — debate engines, alignment runtimes in Rust with zero-knowledge proofs, embedding-based routers, RLAIF loops that generate their own training data. AI that governs itself.

3. Empirical safety research — 54 technical reports measuring what actually happens to safety when you quantize, batch, swap backends, scale concurrency, or change KV-cache precision. Findings backed by TOST equivalence testing, effect-size analysis, and Holm-Bonferroni correction. Three hypotheses overturned. RTSI-gated routing recovers 76% of the quantization refusal gap. One paper accepted at ICML 2026 Hypothesis Testing workshop; five under peer review.

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

Adaptive constitutional engine. Rust alignment runtime. One obsession: make AI systems that are fast, safe, and honest.

Public verification surfaces: Chimeraforge · Chimeradroid · Banterblogs · Echo · JARVIS Console · ProjectWyvern. Research substrate (Banterpacks, Banterhearts, Muse Protocol) is private during the publication window — read access on request via Reach Me.

The core. Everything else in the ecosystem feeds into or out of this.

• Constitutional Debate Engine (TDD001): Multi-model debate with heat-based escalation and 3 consensus algorithms. Constitutional principles as first-class constraints, not afterthoughts.
• Fast-Path Router (TDD002): Embedding-based cosine similarity routing. 99% of queries resolved in <10ms without touching the debate engine.
• Rust Alignment Runtime (TDD005): 7 crates. BFT consensus, Ed25519 provenance chains, zero-knowledge proofs via Pedersen commitments on Ristretto255, CRDT sync for cross-device state.
• RLAIF Pipeline: Debate outcomes generate DPO training pairs that continuously refine the router's alignment centroid. The system improves itself.
• JARVIS Gateway: Unified AI assistant with chat (turn-based state machine), voice (Whisper STT, TTS), graph memory backed by PostgreSQL/pgvector (~40-table cutover from aiosqlite on a dual-dialect SQLAlchemy 2.0 async engine, unified under Alembic), tool execution with human-in-the-loop approval, WebSocket streaming, durable workflows, and a React/Next.js operator console.

The alignment layer doesn't just steer the model. It proves it steered correctly.

The measurement engine. Every claim in the research program comes from code running here.

• Capability-aware backend dispatching — Ollama, HuggingFace Transformers, ONNX Runtime, TensorRT. The system picks the right backend for the job.
• Compile+quantize pipeline with latency/accuracy guardrails
• Thompson Sampling auto-optimizer for configuration discovery
• GPU monitoring — 100ms power polling, thermal safety, VRAM fragmentation tracking via pynvml
• TensorRT engine building, ONNX model export, torch.compile with Inductor backend
• KV-cache analysis — theoretical + empirical measurement, CUDA graph crash reproduction

If you can't measure it, you can't optimize it. If you can't reproduce the measurement, you didn't measure it.

The tool that ships the research.

Published on PyPI —pip install chimeraforge

• 7 predictive models (R² > 0.85 throughput, > 0.96 VRAM, <1s runtime, zero GPU required) — including an opt-in safety gate sourced from refusal data (v0.3.0), a live chimeraforge safety refusal-screen command against running Ollama models (v0.4.0), and a model-agnostic planner with measure-on-demand (v0.5.0: plan --model <id> accepts any registry/Ollama/HF model, plus suggest/catalog/measure commands; fixes two planner correctness bugs), plus continuous batching + TTFT/TPOT latency modeling + Pareto planning (v0.6.0, 483 tests)
• Dual-language benchmarking harnesses (Python + Rust)
• Python 3.10–3.14 CI matrix with trusted-publishing on tag push
• Operationalizes findings from 54 technical reports into deployment decisions

Research that stays in a PDF is a hobby. Research that ships as a CLI is engineering.

Unity/C# JARVIS client for Android and Android XR. WebSocket streaming, voice interface, tool approval UI, CRDT cross-device sync, session handoff. Runs on Android XR headsets for embodied agent work. Phone talks directly to your local laptop GPU — no cloud dependency.

The ecosystem runs everywhere, not just on a dev machine.

The messaging bridge between external platforms and the JARVIS gateway. 5 adapters (Slack Socket Mode, Discord Gateway, Telegram long-poll, WhatsApp Cloud API webhooks, SMTP/IMAP email). Each adapter is a thin HTTP relay — no intelligence, just platform-specific formatting. All cognition lives in JARVIS.

Next.js 15 + React 19 operator surface for JARVIS. Radix primitives, Framer Motion, Tailwind. Chat, tool-approval workflows, session telemetry, and live agent state.

The autonomy plane in the Chimera ecosystem. Sits between the Chimera control plane (identity, policy, operator approvals) and the flight controller (PX4/ArduPilot via MAVSDK/MAVLink/ROS 2). Owns mission validation, command arbitration, telemetry normalization, and cryptographically replayable mission archives.

AI assists planning. It never bypasses deterministic safety or operator authority.

54 technical reports (TR 108–TR 168). ~1.3M decision-grade measurements (curated from ~10⁹ profiler samples). 3 hypotheses overturned. 1 ICML 2026 workshop paper accepted + 5 under peer review.

Audit path: every TR is a markdown file in /reports/ — count, read, diff. No site, no slides, no PDF wall. The folder is the source of truth.

Decision-grade statistical validation: TOST equivalence testing, Cohen's d effect sizes, Holm-Bonferroni correction, bootstrap confidence intervals.

Quantified the safety tax of inference optimization across 4 model families:

Key finding: backend matters more than numerical precision for safety. A 23pp safety drop traced to chat template divergence, not FP16 vs Q4 arithmetic.

First mitigation (TR163): RTSI-gated routing recovers ~76% of the weight-quantization refusal gap by routing the riskiest 20% of configs to direct safety testing. LOOCV ROC-AUC 0.84, validated across LOOCV passes during the QuantSafe Certifier buildout; companion arXiv preprint routes 10/10 hidden-danger configs, Wilson 95% CI lower-bound 0.72.

Proved via Nsight Systems kernel tracing that the multi-agent scaling bottleneck is GPU memory bandwidth physics, not serving software. Continuous batching (vLLM/TGI) amortizes this:

^{Nsight Compute trace from the kernel-profiling pass that produced the measurements above.}

• Multi-agent scaling follows Amdahl's Law (R² > 0.97), throughput plateaus at N=2
• Q4_K_M is the universal quantization sweet spot (30–67% cost savings)
• VRAM spillover causes 25–105x latency cliffs — the real context-length bottleneck, not quadratic attention

1. M/D/1 queueing theory — deviates 20.4x from observed behavior (TR 128)
2. NUM_PARALLEL enables concurrent GPU inference — confirmed no-op, 0/30 tests significant (TR 128)
3. Serving stack is the scaling bottleneck — GPU memory bandwidth physics dominates; PyTorch Direct degrades worse than Ollama (TR 131)

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.

• 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
• Input guardrails across all APIs and agents with adversarial attempt logging
• 5 specialized PR-review agents distilled from ~2,500 comments across ~1,000 PRs
• 5000+ tests across ~20 services

Oct 2025 – Present · New York, USA

Multi-service clinical AI platform for psychotherapy training and research workflows.

• Real-time streaming agent via LiveKit SDK + Google Gemini Realtime API for low-latency WebRTC voice/avatar sessions
• Versioned persona engine, instructor workflows, and Claude-powered clinical evaluation service
• Tiered consent, pattern-based PII scrubbing, research exports, and HIPAA BAAs executed across Anthropic and AWS

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

ML/AI: PyTorch, TensorFlow/Keras, Transformers, DeepSpeed, Accelerate, Ray, RAG, LangGraph, LangSmith, MCP

Web Frameworks: FastAPI, Next.js, React

Inference & Serving: vLLM, SGLang, TGI, TensorRT-LLM, llama.cpp (GGUF), continuous batching, KV-cache optimization, speculative decoding

GPU & Compilation: CUDA, Triton, TensorRT, FlashAttention, ONNX Runtime, torch.compile, Nsight Systems / Nsight Compute, quantization (GPTQ, AWQ, INT4/INT8)

Post-Training: LoRA/QLoRA, RLHF, RLAIF, DPO/ORPO/KTO, GRPO-family RL, policy optimization, PRM/ORM routing, WARM judges

Data & Analysis: PostgreSQL, DynamoDB, Redis, ClickHouse, Qdrant, SciPy, SHAP

Cloud & Deployment: AWS, Azure, Docker, Kubernetes, Vercel

Monitoring: Prometheus, Grafana, Datadog, OpenTelemetry, pynvml, MLflow, W&B

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

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

Medical AI Imaging — Multi-Phase Clinical Pipeline Led a 4-person engineering + clinical team (3 engineers, 1 physician) across a 5-institution program (state government, city university, dental hospital, 2 engineering colleges) building TensorFlow/Keras pipelines over clinical imaging — dental (POC) → retinal (Phase 2) → EEG (Phase 4+). Stack: LSTM + attention multi-classification, W&B experiment tracking, SHAP interpretability, pinned-memory CPU↔GPU transfer optimization.

Registered work: Copyright L-122721/2023.