AI/ML Engineering Manager · Architect and Builder

Building and scaling trustworthy, production-grade AI/ML products — AI/ML, Agentic AI, LLMs, LLM safety & guardrails, evaluation systems, and scalable ML infrastructure (MLOps / AgentOps).

🔭 Currently: how to build, evaluate, calibrate, scale, and safely deploy LLM applications and AI agents efficiently. 💞️ Open to collaborating on open-source in AI/ML, LLMs · Generative AI · agentic workflows · AI evaluation · AgenticOps.

📫 GitHub · Pronouns: He/Him

An MLflow-integrated control plane for trusted agentic workflows: governed prompts, tools, skills, MCP servers, workflows, knowledge bases, evaluations, calibration, observability, deployment, and the Aria copilot. A new Workspace layer adds multi-project/tenant authorization scoping, a pluggable source-control provider contract with a GitHub adapter, a PR-like Change Request review/promotion lifecycle, and release-governance state machines with break-glass policy. The current release tracks MLflow 3.15 and PostgreSQL 17 with MCP 2.x, a layered architecture, 16 UI cookbooks, a published documentation site, and CI-gated UI test coverage above 90%. AI agent platform · MLflow · AgentOps · LLMOps · MCP · RAG · governance · multi-tenancy · evaluation · observability · prompt engineering

Safety benchmark gains do not guarantee safety transfer. A paired, same-checkpoint study of how compact prompt-safety guards specialize, transfer, and compose — LoRA-SFT vs. base-anchored KL-SFT across four instruction checkpoints, a dual-labeled mortgage benchmark, and an analysis-preregistered panel of released vendor guards. The v0.0.1 research hub brings together the published HTML report and the unified report PDF. The release records explicit study state, evidence tiers, verification paths, and redistribution decisions; 28 of 32 generated inputs verify in the standard environment, with the remaining four requiring the pinned analysis environment. LLM safety · AI safety · guardrails · jailbreak detection · prompt injection · content moderation · LoRA / SFT · small language models · reproducible research

Traces establish recurrence, not admissibility. A trace-to-program compiler that turns repeated read-only agent prefixes into deterministic guarded programs — and refuses whenever the evidence cannot license one. Typed value provenance, effect and position barriers, a bounded 23-operator DSL, runtime verification, and a finite-sample selective-risk gate whose default output is retirement. Across three live-provider GitHub workflow families it matches the baseline on 90/90 exact outcomes (versus 89/90), while reducing provider requests by 66.6%; on NESTFUL and API-Bank every recurrent family retires — which is the result, not a failure. Now finalized through an ICLR reviewer-response revision and proofreading pass, the published artifact shelf documents the evidence and its limits. LLM agents · agent optimization · program synthesis · compiler · provenance · tool use · trace analysis · selective risk control · reproducible research

Faithful ICLR 2026 implementation — evolving, self-improving context playbooks for LLM agents via a Generator → Reflector → Curator loop with incremental delta updates. OpenAI Agents SDK support, 163 tests, and an 11-recipe cookbook. context engineering · self-improving agents · in-context learning · agent memory · LLM agents · prompt optimization · RAG · OpenAI Agents SDK

A clean reference implementation of the Agent-to-Agent (A2A) protocol — specialized AI agents that discover each other and collaborate over JSON-RPC 2.0. Python · FastAPI · Pydantic, with 147 tests at 93% coverage and a worked cookbook. A2A protocol · multi-agent systems · agent interoperability · agent communication · JSON-RPC · FastAPI · Pydantic · LLM agents

Enterprise compliance automation — turn compliance documents into queryable knowledge graphs via a 15-agent pipeline (11 extraction + 4 comparison), with executable-readiness gating, independent grounding verification, a coverage-checked dependency-DAG generator, and an interactive graph explorer. knowledge graphs · compliance automation · RegTech · multi-agent AI · grounding verification · JanusGraph · FastAPI · LLM

A focused research spinoff that compiles compliance policy text into typed, source-grounded executable decision logic — DMN/BPMN/CMMN/SBVR plus a LinkML business information model. Its thirteen-agent CLI pipeline independently re-derives each rule's evidence from the raw corpus rather than trusting the citation the rule carries, then partitions the survivors into dependency DAGs with a checked coverage guarantee. RegDelta diffs two versions of a policy and propagates the impact, now layering an optional LLM-assisted semantic comparison over its deterministic diff — scoring unmatched rules for equivalence and routing contradictions to human review rather than auto-resolving them; six properties of the type lattice, bounded prover, and dependency partition are discharged by exhaustive enumeration rather than by example. compliance automation · business rules · decision automation · DMN / BPMN · knowledge graphs · LinkML · grounding · dependency DAGs · differential execution · RegTech · semantic comparison

Reinforcement learning (PPO/A2C/DQN) that dynamically tunes AML risk-scoring weights per case — Gymnasium env, FastAPI backend, React training dashboard. reinforcement learning · anti-money laundering · RegTech · PPO · Gymnasium · risk scoring · FastAPI · React

Domain-agnostic, single-node, docker compose-portable tabular AutoML pipeline (Dagster + FLAML + MLflow + FastAPI) with drift monitoring, an online feature store, a Streamlit dashboard, and one-command start/stop. The real-estate buyer-stage classifier ships as the worked example — all domain knowledge lives in YAML, never in framework code. AutoML · MLflow · Dagster · FLAML · drift detection · tabular ML · feature store · real estate ML

AI-powered mock-interview assistant for ML engineering, leadership/behavioural, and coding interviews. Gradio 6 · LangChain 1.x · OpenAI · Whisper. interview prep · LangChain · GPT · OpenAI · Whisper · speech-to-text · generative AI · Gradio

Agentic AI · LLM safety & guardrails · prompt-injection / jailbreak detection · LLM & agent evaluation · program synthesis & selective risk control · MLflow / LLMOps / MLOps · RAG · reinforcement learning (RLHF/GRPO/PPO) · fine-tuning (SFT/LoRA) · knowledge graphs · multi-agent systems

Python · PyTorch · Transformers · TRL · MLflow · FastAPI · LangChain · React

_{⭐ If any of these are useful, a star helps others find them too.}