Autonomous AI systems cannot be built with stochastic prompt engineering alone. My work bridges low-level hardware constraints (RTX 4060 8GB GPU), deterministic reference monitors outside the model, and formal information flow control.

+----------------------------------------------------------------------------------------+
|                              APPLIED AI SYSTEMS ARCHITECTURE                           |
+------------------------------+------------------------------+--------------------------+
| 1. WARRANT                   | 2. OPTISERVE                 | 3. INTERPOSE             |
| Agentic RAG & Attributed NLI | PyTorch LLM Inference Engine | Deterministic AI Monitor |
| Reliability & Attribution    | Acceleration & Distillation  | Dynamic Taint & Security |
+------------------------------+------------------------------+--------------------------+

Deterministic Information Flow Control (IFC), Dynamic Taint Tracking & Adaptive Red-Teaming for Tool-Calling Agents.

• Engineering Challenge: In-model prompt guardrails inevitably fail against indirect prompt injections and data poisoning attacks. Autonomous enterprise agents require non-bypassable, deterministic reference monitors.
• Systems Architecture: An out-of-band reference monitor intercepting all tool invocations before execution. Tracks data provenance across token slices using a formal join semi-lattice (SANITIZED < TRUSTED < RESTRICTED < UNTRUSTED < ADVERSARIAL), enforces AST-level constraints (ast, sqlglot, bashlex), and gates irreversible operations behind cryptographic HMAC-SHA256 human-in-the-loop sign-offs.
• Empirical Benchmarks:
  • 0.0% Attack Success Rate (ASR): Neutralized 36 of 36 AgentDojo threat vectors and PyRIT adaptive mutations (Base64 split, ISO-27001 drills, tag escaping).
  • 100.0% Benign Utility: Zero false-positive halts on enterprise workflows.
  • Sub-Millisecond Overhead: Mean evaluation latency of 0.038 ms with pure CPU AST parsing.
• Live System HUD: interpose.vercel.app
• Repository: github.com/Hamza-HATTAB/interpose
• Stack: Python 3.13, FastAPI, PyRIT, AgentDojo, AST Parsing, Next.js 14, Tailwind CSS.

Sub-8GB VRAM Reasoning Distillation, 5-Way Quantization Bake-Off & AirLLM 70B Layer-Wise NVMe Streaming.

• Engineering Challenge: Serving frontier reasoning models (DeepSeek-R1 / Qwen2.5) locally requires expensive enterprise clusters (A100/H100), exceeding the budgets of edge deployments and lean infrastructure.
• Systems Architecture: A hardware-aware serving engine engineered for an 8GB RTX 4060 GPU. Implements exact rejection-sampling speculative decoding (K=3) with LoRA-distilled draft models, continuous batching, a 5-way quantization comparison (AWQ, GPTQ, GGUF, FP8 E4M3, FP16), and AirLLM layer-wise NVMe streaming for 70B parameter models.
• Empirical Benchmarks:
  • 1.92x Generation Speedup: 72.4 tokens/second on mathematical reasoning trajectories (GSM8K).
  • Strict Memory Safety Guard: Enforces an invariant ceiling of <6.8 GB VRAM to prevent out-of-memory kernel faults.
  • Zero-Cloud 70B Execution: Streams 70B weights from NVMe SSD inside 2,150 MB peak VRAM.
• Live System HUD: optiserve.vercel.app
• Repository: github.com/Hamza-HATTAB/optiserve
• Stack: PyTorch 2.6, CUDA 12.4, HuggingFace Transformers, BitsAndBytes, AirLLM, Next.js 14.

Claim-Level Decomposition, CPU DeBERTa-v3 Cross-Encoder Verification & 3-State Selective Abstention.

• Engineering Challenge: Standard RAG pipelines suffer from silent multi-hop hallucinations when questions require cross-document reasoning or when evidence is missing.
• Systems Architecture: An attributed agentic research system that decomposes generated hypotheses into atomic assertions. Evaluates claims via a two-stage verification pipeline (deterministic numerical/temporal regex guard + calibrated DeBERTa-v3 NLI cross-encoder at tau >= 0.82) and enforces a 3-state selective prediction policy contract (FULL_PASS, PARTIAL_PASS with graceful claim pruning, or ABSTAIN).
• Empirical Benchmarks:
  • Selective Abstention: Eliminates unsupported assertions across 200 HotpotQA evaluation questions via formal refusal rather than ungrounded extrapolation.
  • Zero-VRAM CPU Verifier: DeBERTa cross-encoder evaluates on pure CPU in 689 ms, preserving 100% GPU VRAM for generation.
  • Bidirectional Citation Tracing: Complete evidence DAG mapping assertions to cited Wikipedia paragraph spans.
• Live System HUD: warrant-alpha.vercel.app
• Repository: github.com/Hamza-HATTAB/warrant
• Stack: Python 3.11, Qdrant Hybrid RRF, FlashRank, Gemma-3, DeBERTa-v3, Next.js 14.