# Vidit Ostwal > ML/AI Engineer building agentic AI systems and RL environments. Site: https://vidit-ostwal.github.io Generated: Jun 21, 2026 --- ## Home URL: https://vidit-ostwal.github.io/ Role: ML / AI Engineer ### My Story I graduated from BITS Pilani in 2023 and have been working as a Gen AI Engineer since — iterating on agents, RAG pipelines, and fine-tuning. Outside of work, I contribute to open source projects. Around January 2025, I started exploring CrewAI. I really wanted to understand the core of how agent orchestration frameworks are built. I went deep — contributed 35+ merged PRs, became the 11th contributor on the open-source repo, and placed 2nd runner-up at their Fall Hackathon. By December 2025, I had enough credibility to join the review team for a book on the framework that's about to be published. After that, my curiosity shifted toward RL environments. I started exploring Prime Intellect and OpenEnv, built a few playful environments, and now I'm excited to work on building environments genuinely useful for training frontier models. I've contributed to CrewAI, Ragas, GitHub MCP, OpenEnv, and Prime Intellect. Check out my GitHub to see the open-source work. I actively share what I am building on X. ### Links - GitHub: https://github.com/Vidit-Ostwal - LinkedIn: https://www.linkedin.com/in/viditostwal/ - X: https://x.com/ViditOstwal - Email: mailto:viditostwal@gmail.com - Anonymous feedback: https://www.admonymous.co/vidit-ostwal ### Open-source contributions - CrewAI: https://github.com/crewAIInc/crewAI - Ragas: https://github.com/vibrantlabsai/ragas - GitHub MCP: https://github.com/github/github-mcp-server - OpenEnv: https://github.com/meta-pytorch/OpenEnv - Prime Intellect: https://github.com/PrimeIntellect-ai/verifiers ## Experience URL: https://vidit-ostwal.github.io/experience ### Machine Learning — Generative AI — Piramal Capital & Housing Finance Limited Period: Nov 2023 — Present (current) - Designed and developed ARYA, an internal chatbot for the sales team using RAG and agentic systems — handled 500,000+ queries since launch and won Best AI Chatbot Product of the Year. - Built Daily-Huddle, a WhatsApp-integrated chatbot for Branch Sales Managers, reducing daily huddle time by 30%. - Designed CAM-Copilot, an underwriter assistant that summarises CIBIL reports, bank statements, and documents — cutting turnaround time by ~40%. - Currently building an internal Text-to-SQL chatbot inspired by Uber's QueryGPT. ### Data Engineer — Piramal Capital & Housing Finance Limited Period: Jul 2023 — Oct 2023 - Orchestrated data pipelines for real-time feature delivery to ML models, improving model performance by 15%. - Designed and deployed scalable microservices architecture, reducing feature serving latency by 20%. - Used Python and Spark to optimise end-to-end feature store functionality for various ML workloads. ### Junior Associate — Indus Insights · Internship Period: Jan 2023 — Jun 2023 - Developed dynamic forecasting models using delinquency buckets and economic adjustments for seasonal and annual client planning. - Led client meetings presenting findings on acquisition, response rates, and intent validation. - Took initiative on ad-hoc analyses beyond core responsibilities, demonstrating adaptability. ### University Leader — Community Classroom Period: Jul 2021 — Sep 2021 - Founded and led a university community promoting accessible education for all. - Conducted workshops to foster a programming culture on campus. ### Project Leader — Machine Learning — Swecha · Internship Period: Jun 2021 — Jul 2021 - Designed and implemented a fact-checking ML model, scraping news articles from reputable sources. - Compared scraped articles against user-input news using textual entailment analysis to enhance accuracy. - Developed skills in machine learning, data scraping, and NLP. ## Projects URL: https://vidit-ostwal.github.io/projects ### Automating 3D Survival Playing (Jun 2026) Autonomous vision-language agent that plays Survival 3D on YouTube Playables — screenshots the game, reads goals from the UI, plans movement, and executes keyboard actions in a loop. GitHub: https://github.com/Vidit-Ostwal/Automating-3DSurvival-Playing Launch post: https://x.com/ViditOstwal/status/2066886438399139953 - Closed-loop agent on YouTube Playables: capture screenshot → read goal, progress, and inventory → plan 1–2 compass moves → send key presses → persist memory for the next cycle. - GoalMakerLLM reads the game UI and picks the fastest heading; PlannerLLM turns that into concrete NORTH/SOUTH/EAST/WEST steps — each component can use OpenAI or local Ollama vision models. - MemoryManager keeps JSON knowledge of mechanics, discoveries, and failed strategies; RunLogger saves per-cycle screenshots and run.json under outputs/. - Playwright + Chrome with a persistent browser profile for one-time YouTube login; uv-based Python 3.11+ setup with configurable GAME_URL and model providers. ### Automatic UI Fixing Harness (Jun 2026) Black-box UI defect discovery for locally hosted web apps — Playwright exploration, LLM-assisted goals and verification, and an HTML findings report. GitHub: https://github.com/Vidit-Ostwal/Automatic-UI-Fixing-Harness - Four-phase pipeline: BFS exploration in a real browser → LLM converts paths into plain-English test goals → parallel goal-driven executors → per-step LLM verifier that emits structured findings. - Executors resolve instructions at runtime instead of replaying raw trajectories — app-agnostic design with Docker lifecycle management, DOM heuristics, and crash-detection fallbacks when the LLM is off. - Single CLI (run_harness.py) with composable modes — full pipeline, planner-only, run-goals, multi-pass rollout, and report-only — default target is a buggy Memos Docker image, but the harness generalizes to any local web app. ### Understanding Prime Env Skill (May 2026) Skill and installer that produce a rich, self-contained HTML report explaining any Prime Intellect verifiers environment. GitHub: https://github.com/Vidit-Ostwal/Understanding-prime-env Live: https://www.npmjs.com/package/understanding-prime-env Launch post: https://x.com/ViditOstwal/status/2058570288993042477 - Walks the environment source and emits environment_overview.html—dataset and task samples, every reward/metric with weights and scoring range, rollout logic from prompt to final score, and a full config reference. - Ships as an npm CLI (npx understanding-prime-env) with interactive setup plus flags for Claude Code, Cursor, Windsurf, GitHub Copilot, Zed, or all tools in one go. - Meant to be used from inside a verifiers env folder so you can ask your assistant to “explain this environment” and get an overview styled to match the Prime Intellect platform. ### Price Negotiation RL (Mar 2026) RL environment where an LLM agent negotiates prices against an LLM-powered seller using real marketplace listings. GitHub: https://github.com/Vidit-Ostwal/price-negotiation-rl-OpenEnv Live: https://viditostwal-price-negotiation.hf.space/ Launch post: https://x.com/ViditOstwal/status/2052078129263571069 - Three difficulty levels (Easy / Medium / Hard) with varying zones of possible agreement — from a $480 ZOPA down to no ZOPA, testing walk-away discipline. - Six-component reward system on a [−1, 1] scale: surplus capture, walk-away correctness, output compliance, closing speed, opening offer quality, and concession smoothness. - Walk-away penalty weighted 5× — trains the agent to value discipline over forcing a bad deal. - Pure Python arithmetic grader ensures reproducibility; stochastic LLM sampling on both sides creates diverse training signal across episodes. - Deployable on Hugging Face Spaces, Docker, or locally via uv; ships with a browser playground showing real-time reward breakdowns. ### Maze Bench (Apr 2026) Ice-sliding multi-player maze environment on OpenEnv for benchmarking planning and coordination in RL agents. GitHub: https://github.com/Vidit-Ostwal/Maze-Bench-OpenEnv Live: https://viditostwal-maze-env.hf.space/web/ Launch post: https://x.com/ViditOstwal/status/2052373348706091170 - Ice-sliding mechanics — agents slide until hitting a wall or another player, requiring multi-step lookahead over reactive movement. - Simultaneous multi-player movement; solved only when every player reaches an exit cell in the same phase. - Reward shaping penalises repeated actions (−1), reversals, and revisited board states scaled by prior visit count. - Standard OpenEnv reset()/step() API — compatible with any RL training loop without environment-specific wrappers. - Includes dataset validation tooling and GIF rendering from recorded rollouts for debugging agent trajectories. ### MakeMyDocsBot (Oct 2025) CrewAI-powered bot that auto-syncs multilingual documentation across feature branches via a Git pre-push hook. GitHub: https://github.com/Vidit-Ostwal/MakeMyDocsBot Launch post: https://x.com/ViditOstwal/status/1984629588887212078 - 2nd Runner-Up at the CrewAI Fall Agentic AI Challenge. - Detects English documentation changes and auto-generates synchronised translations (Korean, Brazilian Portuguese) with no manual intervention. - Runs as a Git pre-push hook — fires before any feature branch push, prompting the developer to sync docs before code lands. - Agent-based architecture with specialised CrewAI agents for change detection, translation, and cross-branch synchronisation. - Minimal setup: Python 3.12+, uv, and an OpenAI key — clone + uv sync + one hook registration. ### RLM Demo (Jan 2026) Full-stack interactive console for demonstrating and chatting with Reinforcement Learning models and agentic systems. GitHub: https://github.com/Vidit-Ostwal/RLM-demo Live: https://huggingface.co/spaces/ViditOstwal/RLM-Interactive-Console Launch post: https://x.com/ViditOstwal/status/2014769757434872304 - Chat interface for real-time interaction with RL models, backed by smolagents for agent-based reasoning and step-by-step trace display. - FastAPI backend (Python 3.12+) paired with a Next.js 15 / React 19 frontend — clean separation of model logic and UI. - Integrates Hugging Face datasets directly; response caching via JSON files keeps repeated queries fast. - Docker-first production deployment — single container exposes the full stack on port 7860, ready for Hugging Face Spaces. - Environment-driven config (HF token, model name) makes swapping the underlying model a one-line change. ## Writing (external) URL: https://vidit-ostwal.github.io/blog - KL Divergence — Made Visual (Jun 12, 2026) — LLM Behavior & Sampling An interactive visual guide to KL divergence: distributions, the formula, live computation, and how it shows up in LLMs and PPO training. /blog/kl-divergence - MoE Routing Calculation (Excel Walkthrough) (Jan 25, 2026) — Mixture of Experts Step-by-step Excel breakdown of router logits → top-k selection → normalized expert probabilities. Includes top-k masking, −∞ replacement, softmax, and final expert routing weights. https://docs.google.com/spreadsheets/d/1jVwHc-BfOCop92g6tAfQOeTDb8Nfl3yunvQ3dwR1gI0/edit?usp=sharing - Building MakeMyDocsBot (Dec 20, 2025) — Agentic Systems & Tooling Automated multi-language documentation sync across feature branches. https://viditostwal.substack.com/p/building-makemydocsbot - How Does Temperature Change LLM Responses? (Jul 9, 2025) — LLM Behavior & Sampling Effect of temperature on next-token probability distribution. https://viditostwal.substack.com/p/how-does-temperature-changes-the - KV (Key-Value) Cache in Transformers (Jul 26, 2025) — Transformers & LLM Internals Reducing inference latency using KV cache. https://viditostwal.substack.com/p/kv-key-value-cache-in-transformers - Masked Self-Attention (Jun 25, 2025) — Transformers & LLM Internals How masking enforces autoregressive generation in decoder-only transformers. https://www.notion.so/viditostwal/Self-Attention-in-Transformers-216e478805d48005b515fac90e1d76e0 - Self-Attention in Transformers (Jun 21, 2025) — Transformers & LLM Internals How queries, keys, and values compute attention weights and why it matters. https://www.notion.so/viditostwal/Self-Attention-in-Transformers-216e478805d48005b515fac90e1d76e0 - Training the Tokenizer (Jun 3, 2025) — Transformers & LLM Internals How tokenizers are trained and why the vocabulary choice shapes model behavior. https://www.notion.so/207e478805d48090b34fcc5c8e8c3c01?v=207e478805d480cfac6c000ca3c80482 ## Site map - Home: https://vidit-ostwal.github.io/ - Experience: https://vidit-ostwal.github.io/experience - Projects: https://vidit-ostwal.github.io/projects - Writing: https://vidit-ostwal.github.io/blog - LLM digest: https://vidit-ostwal.github.io/llms.txt