AI research + knowledge systems

Marvin B. Freedman

AI Research Engineer — LLM knowledge, retrieval, and verification systems.

I build practical AI systems that help language models retrieve, remember, verify, and act on external knowledge. My work spans agent workflows, RAG-style retrieval, long-context memory, AI coding assistants, product chatbots, verification loops, and full-stack applications that turn ambiguous AI ideas into working systems. I study how language models fail when interacting with external knowledge — in retrieval ranking, context boundaries, verification signals, and long-horizon memory — and build systems to measure and fix those failure modes.

Featured project GitHub profile

Anthropic Knowledge Team fit

External knowledge, retrieval, agentic search, and model reliability.

My strongest fit is building systems around LLMs that make external knowledge usable: scoped memory, retrieval workflows, verification loops, source-of-truth files, catalog/search-backed chatbots, and agent operating patterns.

AiML SuperAgent is the strongest signal: it treats AI agents as long-running operators that need memory boundaries, context discipline, verification checks, audit trails, and repeatable handoffs.

The surrounding work connects that model to practical knowledge systems: RAG-style product search, chatbot memory, AI-assisted coding workflows, technical writing on verification, finetuning dynamics, RL-style feedback loops, distillation, and full-stack AI products.

I am interested in how frontier models should search, retrieve, rank, remember, verify, and reason over information, including indexing, query understanding, knowledge graphs, and distributed retrieval systems.

First featured project

AiML SuperAgent

Agent memory + verification framework

Agent memory + verification framework for long-running AI coding systems.

AiML SuperAgent turns AI coding work into a disciplined operating loop: scoped project memory, source-of-truth files, working notes, verification checks, deployment logs, safe environment audits, and repeatable handoffs across sessions.

Visit AiML SuperAgent View GitHub repo

Context discipline Keeps durable memory separate from active working context so agents load only what matters.
Verification loop Pushes AI work through concrete checks before notes, deployments, or claims are treated as true.
Project operations Connects code changes, deployment notes, incidents, and safe-env handling into one repeatable workflow.

Retrieval, verification, agent workflows, and applied knowledge systems.

AI-assisted verification

RH Verification System with machine-checked numerical certificates and reproducible Python pipelines

Built an AI-assisted Python pipeline that systematically tested 48,000+ mathematical reductions, closing 293/293 formal theorem certificates with machine-verified numerical bounds. Companion to an RH Weyl-positive verification preprint and public reproducibility supplement.

Open the visual explainer

Finetuning + RL research note

Verifier-Calibrated On-Policy Distillation for feedback-driven model behavior

Practical research note on finetuning dynamics: how verification signals, RL-style on-policy feedback loops, and dense distillation targets affect model behavior, and how to preserve useful capabilities.

Read the VC-OPD article

AI commerce assistant

AI sales chatbot with catalog retrieval, product recommendations, policy-aware answers, and human handoff

SellBotAiML is a Shopify-first AI sales assistant that uses synced product catalogs, approved store policies, multilingual chat, product recommendations, and operator handoff workflows to help shoppers find the right products and move toward checkout.

Visit SellBotAiML.com

Catalog retrieval

Catalog-backed product knowledge chatbot for ecommerce retrieval and customer support

SchoolRugs is a live product knowledge surface where catalog-backed retrieval, product search, chatbot memory, and customer support workflows help buyers find relevant rugs and answers.

Visit SchoolRugs.com

Agent systems

Agent operating files, durable context, and reliable AI coding workflows

A public article reference on why a compact CLAUDE.md-style operating file matters for reliable coding agents, durable context, handoffs, and repeatable AI-assisted engineering behavior.

Open the Medium article