Arash Farahani, Ph.D.

Clinical AI & Knowledge Graphs

• •Clinician-in-the-Loop Self-Learning Loops
• •MedKG Auto-Generated Graphs (Provenance)
• •Factor-to-NumPyro Compilation & Bayesian Inference
• •Value of Information (VOI) Action Ranking
• •Medical Safety & Harm Evaluation (NoHarm/MAST)
• •Clinical Reasoning Engine Parity

Post-Training & Agents

• •SFT & Instruction Tuning
• •PEFT Benchmarking (LoRA, DoRA, AdaLoRA, Full)
• •Hyperparameter Optimization & LoRAX
• •Agent Harness & Programmatic Verifiers
• •Reward Signal Design (Unlabeled Telemetry)
• •In-Context Self-Correction (Reflexion-Style)

LLM Systems & Inference

• •Agentic Workflows & Tool-Calling
• •Dynamic Multi-Vector RAG Engine
• •In-Process Graph-Grounded LLM Judges
• •Inference Hallucination Detection
• •CUDA Kernel Development (int8 GEMM + fp16 LoRA)
• •Multi-Step Reasoning & Closed-Loop Correction

Infrastructure & Tooling

• •AWS Bedrock, SageMaker, Fargate, Lambda
• •AWS CDK (IaC) & Production CI/CD
• •NumPyro & Probabilistic Programming
• •Apache Spark (Scala)
• •Distributed Training Pipelines
• •GraphQL, RPC & Chrome Extension APIs

Languages & Frameworks

• •Python (Expert)
• •PyTorch
• •TypeScript & Vue.js
• •C++ / CUDA
• •SQL
• •Stata & R

• •Architected a closed-loop clinician-in-the-loop self-learning pipeline: automatically ingested attending physician feedback and clinical corrections, classified failure modes via an automated Opus 5.5 evaluation judge into structured clinical safety taxonomies, and programmatically generated regression test suites that drove automated prompt optimization rounds.
• •Co-architected MedKG, an auto-generated probabilistic medical knowledge graph with strict citation provenance: extracted authoritative clinical evidence (First Databank, HPO, clinical society guidelines) into grounded factor graphs compiled to NumPyro for Bayesian posterior updates and Value of Information (VOI) calculations to rank next diagnostic actions; built an in-process graph-grounded judge (ADR-034) cutting guideline evaluation latency by ~60% with zero hallucinated guideline violations.
• •Led the clinical AI safety and harm-reduction evaluation programme: lifted benchmark diagnostic F1 from 0.881 to 0.901 across standardized clinical inputs (surpassing published state-of-the-art leaderboards), cut harmful clinical verdicts by 5.8 percentage points, and reduced severe risk rates to 0.045.
• •Increased Clinical Guideline (COG) citation grounding from 38% to 49% by quoting retrieved clinical evidence passages directly into physician consult drafts, driving guideline contradiction rates below 2% and resolving 35 must-do guideline coverage gaps across the US medical corpus.
• •Engineered production Sol parity on Tier 2 clinical reasoning engines: reached 94.0% Overall Accuracy and 95.8% PPV — matching frontier Claude 3.5 Opus reference models while reducing adverse approval errors across 2,000+ patient visits; validated safety mechanisms via rigorous leave-one-out ablations (bipolar stop saving 23 bad approvals, override narrowing fixing 6 cases for 0 broken).

• •Designed and built a closed-loop agentic runtime: LLM generates JavaScript from natural language, injects it into a live DOM, evaluates execution success via a visual diff gate and console error taxonomy, and routes failures to targeted fixers — a self-healing verbal-reinforcement loop (Reflexion-style) driven by a programmatic verifier.
• •Built a reward signal pipeline without labeled data: structural signals (selector rank, turn efficiency), outcome signals (DOM diff, event log intercept), and behavioral signals (reset rate as implicit failure label) — each tier calibrated to cost and latency.
• •Engineered DOM compression reducing ~50K-token page context to ~800 tokens via a semantic clustering algorithm with stable cross-render identifiers — enabling practical per-action LLM calls.
• •Designed a three-vector RAG engine (prior scripts, extracted selectors, failure patterns) that self-updates from every execution, improving generation quality without retraining.
• •Built and maintained the full frontend: Chrome extension UI and shared frontend components in TypeScript/Vue.js (~70K lines authored solo across both repos) — sidebar chat interface, script injection runtime, auth flow, and real-time sync with the backend.
• •Shipped to production on AWS (Fargate, Lambda, CDK): GraphQL backend, vector database, authentication, payments, subscriptions, real-time script sync.

• •Co-led applied research for CodeWhisperer enterprise code customization from inception to General Availability — designed the approach for adapting a code LLM to private enterprise codebases (RAG + LoRAX-based fine-tuning, per-customer adapters on a shared base), serving millions of developers.
• •Led the fine-tuning research that defined CodeWhisperer’s production recipes: systematic sweeps across all LoRA hyperparameters (rank, alpha, target modules, dropout, learning-rate schedules) and controlled comparisons of LoRA vs. full fine-tuning, DoRA, and AdaLoRA on code-generation quality and training stability.
• •Investigated inference-cost reduction for multi-tenant adapter serving — began developing a custom CUDA kernel computing unmerged LoRA paths over an int8-quantized base model, fusing the quantized base GEMM with the fp16 low-rank path.
• •Studied hallucination failure modes in code LLMs: developed an inference-time detector for hallucinated symbols (nonexistent functions, methods, classes) that blocked invalid generations in production; shipped an import-adder system for context-aware completion.
• •Developed evaluation methodology grounding model quality in real-world outcomes rather than benchmark proxies: execution pass rate, symbol-hallucination rate, developer acceptance rate.
• •Supervised intern research on LLM-based LLVM compiler optimization, directing model fine-tuning and evaluation.

• •Engineered large-scale financial data pipelines in Apache Spark (Scala) for entity resolution across millions of records.
• •Founded "Clean Coders" engineering guild, establishing org-wide standards for test-driven development and software craftsmanship.

• •NYC Independent Budget Office, Economist (2017–2019): Led quasi-experimental causal inference research — including a difference-in-differences evaluation of a business incentive program — and communicated findings to non-technical policymakers, including expert testimony before the NYC City Council.
• •Ph.D. research (2010–2017): Applied instrumental variables, time series forecasting, and regression discontinuity design to public policy and infrastructure economics questions.