Nadeem Khan

Projects

Things you can run in a browser first, then the code behind them.

Pinned

  • nl2sql playground: the Ask tab with the search index, the sample database and guided questions

    nl2sql playground

    Ask your database questions in English. The model emits a typed query plan, never SQL text - validated against the real schema and the caller's role before any SQL is generated.

    question → planner (typed plan, never SQL) → validator (schema, joins, role) → sqlglot → read-only executor

    A plan that fails validation never becomes SQL, and a security refusal is final. On PyPI as nl2sql-engine.

    Open demo (opens in a new tab)Code (opens in a new tab)

  • Post-training docs site: the overview page with scope, quickstart and tests

    Post-training

    A collection of small, runnable implementations of LLM post-training and alignment methods, from RL basics to DPO, RLHF, and RLAIF.

    environment → agent (Q-learning, DQN, PPO, GRPO) → policy update → tests

    Each method is small enough to read in one sitting, and each has its own tests.

    Read the docs (opens in a new tab)Code (opens in a new tab)

  • Decision Arena: two decision models set up side by side on the highway scenario

    Decision Arena

    Decision Arena: TypeSafe's Jev vs open-source Laya playing highway-env, Snake and Blackjack with zero training, plus benchmarks and a Claude Code watchdog

    game state → decision model (Jev or Laya) → probability per allowed action → environment step

    The model can only answer with an option on the list. Includes agent-watchdog, which scores each Claude Code tool call before it runs.

    Open arena (opens in a new tab)Code (opens in a new tab)

  • RAG playground: the index pipeline steps beside the Ask panel and its retrieval settings

    RAG playground

    A local-first bench for learning and demonstrating RAG by experiment

    PDF → parse → clean → chunk → index → retrieve (dense, keyword or hybrid) → rerank → cited answer

    Every stage is swappable, and evaluation says why each miss missed.

    Open demo (opens in a new tab)Code (opens in a new tab)

  • logscribe (opens in a new tab)

    AI-powered log analysis for Python logging: batch, scrub PII, and route logs to an LLM for insights.

    Python, Aug 2026

  • medalflow (opens in a new tab)

    dbt, but in Python classes. Declare medallion (Bronze/Silver/Gold) models as Python classes; MedalFlow extracts dependencies from your SQL and compiles them into a staged execution plan.

    Python, Aug 2026

Other repositories

Earlier work, before 2024