Working as an AI Engineer at Square Yards(Nov 2025 — Present), I've built custom MCP servers to integrate with AI apps and also iterating on voice agents’ post-call harnesses to reliably identify and qualify leads among different business verticals to bring down false positives.
Mumbai, India
Things I’ve worked on
Improving Agent using GEPA
Built an enterprise SaaS support and identity environment to evaluate and improve an AI agent using GEPA (Genetic-Pareto Reflective Prompt Evolution) optimization. Across 8 hard cases, the baseline agent achieved a 25% pass rate (2/8 tasks) and an average task score of 0.8875. After three GEPA optimizer iterations, it achieved a 100% pass rate and an average task score of 1.0. Minimax M3 powered the agent, while GLM-5.2 served as GEPA's teacher model.
Built a small decoder-only transformer from scratch to learn all the nuances of model building, reached an exhaustive accuracy of 99.78% for arithmetic operations, also post-trained it using SFT and explored RL using GRPO for improving existing capability. Concluded with data/representation mattered more in pre-train than RL unless the model is already good.
Built a shared context cloud for agents such as Claude Code and Codex, helping siloed teams and isolated machines share project context across teams to accelerate development. It is organization-scoped, project-isolated, and exposed through MCP, making it easy to plug into any compatible agent.
Built a Python SDK for auditing and governing AI agent runtimes by capturing execution events, enforcing policies and guardrails, redacting sensitive data, and generating tamper-evident evidence packs with hash-linked integrity for compliance review and operational oversight. Open-sourced it with Apache-2.0 license.
Built a full-stack local AI agent builder in Go + Next.js with SQLite-backed agent specs, a shared Go run engine, SSE chat streaming, a tool registry with approval gates, edit-agent chats, source tracking, version snapshots, and OpenAI-powered custom agents.
Built infrastructure for reliable agentic Office document editing by converting PPTX/OOXML files into canonical graph representations and applying deterministic document mutations. Solves the reliability gap in automated white-collar document workflows through structure-aware transformations, layout-preserving edits, and render validation for long-tail document operations where current AI fails (last 10% editing) to perform actions accurately.
Built a Python coding assistant with sandboxed shell execution, file tools, streaming OpenAI responses, persistent sessions, and path validation for safer agent workflows.
Built an experimental memory framework for LLM agents to store, retrieve, decay, and reason over long-term memories using SQLite-backed local storage. Added cleanup logic that can rewrite, invalidate, and remove stale memories to keep context useful and redundancy-free.
Built a from-scratch TypeScript deep-research agent directly on the OpenAI raw HTTP API, without the Agents or Assistants SDK, to own the while-loop harness end to end. The LLM supplies intent while deterministic code handles Zod argument validation, tool dispatch, recursion, persistence, isolated recursive sub-agents, Exa-powered web search and reading, and a live terminal UI that streams cited answers.
GitHub