Apr 2025 – Present · 1 yr · Surat, Gujarat, India

• Own the AI systems architecture layer for Rocket.new ($24.5M from Salesforce Ventures + Accel). Era 3: full agentic orchestration. Non-deterministic at every layer.
• Designed multi-agent orchestration: specialized agents for research, UI, backend, deployment from one prompt. Stateless agents with externalized context store. Failure recovery, retry logic, inter-agent context passing without human intervention.
• Built hybrid agent architecture to minimize cascading failures. Utility function scores output quality per stage. RL signals adjust routing weights when model-task combinations underperform.
• Multi-model routing across Anthropic, OpenAI, Gemini. Routing based on task complexity, context window, determinism needs, and cost. A/B tested on real user traffic. Model evaluation with automated scoring: accuracy, latency, hallucination rates, cost per task.
• Agents use tool calling, multi-step planning, memory (short + long term), self-correction loops, and output guardrails for hallucination and safety.
• Multi-tenant platform serving 400K+ developers. Rate limiting, per-user quota management, API cost controls. Monitor latency, agent failure rates, cross-agent context loss as core SLIs driving weekly architecture decisions.

Oct 2023 – Jul 2025 · 1 yr 9 mos · Surat, Gujarat, India

• Era 2: AI-augmented generation. The company pivoted from deterministic rule-based code gen to LLM-integrated pipelines. Launched WiseGPT. Grew from 125K to 200K+ active users. I moved from algorithmic feature work to system architecture, bridging the old deterministic world with the new probabilistic one.
• Built WiseGPT's codebase integration: local project analysis, dependency mapping, context-aware code generation inside VS Code. Designed the RAG architecture: embedding-based code retrieval with hybrid search (vector + keyword), reranking for relevance, chunk size optimization to maximize context quality within model windows. Measured retrieval quality systematically to ensure the right code was surfacing.
• Ran and evaluated open-source models locally (Llama, Mistral) alongside commercial APIs. Fine-tuned models on domain-specific training data for code generation tasks. This hands-on experience with model internals informed the routing and evaluation decisions in the production pipeline.
• Owned the Kotlin backend for multi-framework generation. The pipeline now had two brains: the original rule-based engine for structural accuracy and the LLM layer for semantic understanding. Making them work together without contradicting each other was the architecture challenge.
• Pushed design-to-code accuracy from ~70% to 85%. The old rule-based system hit a wall at 70%. The LLM integration broke through it but introduced non-determinism. The last 15% was all edge cases: nested components, responsive breakpoints, framework-specific idioms where the AI and the rules disagreed.

Oct 2022 – Oct 2023 · 1 yr · Surat, Gujarat, India

• Post-Series-A ($9.5M). This role straddled two architectural eras: finishing the hardest problems in rule-based generation while the team began integrating LLMs. Moved from feature contributor to module owner.
• Built the component detection system from scratch. Not just Figma API traversal. ML-based image recognition that could analyze raw screenshots and identify UI components (buttons, cards, lists, navbars) without human labeling. Recursive parsers for nested Figma layer structures. Layout algorithms for flex detection, grid detection, absolute vs relative positioning. Constraint solving for responsive layout generation. This system processed millions of design files in production.
• Owned the template-driven code synthesis layer and the framework-switching logic. Same design input generating idiomatic code across Kotlin, Flutter, and React. Each framework has different component models, state management patterns, and layout paradigms. The bar: generated code had to be indistinguishable from human-written code.
• Built the test suite for verifying generated code correctness. Tracked accuracy metrics systematically. Optimized generation speed. Built specific solutions for the edge cases that broke the pipeline at scale.
• Built REST APIs in Kotlin for the full design-to-code path: design file parsing, component detection, layout analysis, code generation. This was the plumbing connecting raw Figma input to production-ready output.
• Product Hunt #1 Product of the Week during this period.

Nov 2021 – Oct 2022 · 11 mos · Surat, Gujarat, India

• Era 1: purely rule-based code generation. No AI. No LLMs. Just algorithms. Founding-era team. I wrote the first versions of the core systems that still power the product today. Built the product version that won Golden Kitty Award for Developer Tool of the Year 2021.
• Wrote the FIRST Figma design parser. The FIRST code generator for Flutter and React. From zero. No existing code to reference. No library to fork. Just the Figma API docs, the framework docs, and the problem.
• Designed APIs that are still in use in the product today. Architecture decisions I made during this period still shape how the platform works.
• Solved the hardest unsolved problem in the pipeline: designers use Figma in wildly different ways. Some nest 50 layers deep. Some use auto-layout. Some use absolute positioning for everything. The parser had to handle all of it and produce clean, usable code regardless.
• Built multi-page and multi-screen app generation with state management code. Not just individual screens. Full apps with navigation, state flow, and data passing between screens.
• Worked on the Figma plugin itself. Built specific features that were key to the Golden Kitty Award and Product Hunt launch. Optimized generation speed for the demo and launch experience.
• Worked around Figma API limitations and undocumented behaviors. Built workarounds that became core parts of the pipeline.
• Joined customer calls directly. Feedback from those calls shaped features I built. This wasn't a PM relaying requirements. I heard the problems and built the solutions.
• Broad scope: frontend, backend, Figma API integration, parsing logic, component detection, code generation templates, Figma plugin. Everyone owned everything.

Aug 2021 – Nov 2021 · 3 mos · Surat, Gujarat, India

• Fastest intern-to-full time conversion at DhiWise. Production codebase from day one. No training program. Ship or leave.
• Built Flutter screens and widgets that shipped to production. Wrote code generation templates that defined how the system produced output. Built internal developer tooling used by the engineering team.
• Fixed critical production bugs. Wrote documentation of the architecture and codebase for engineers who would join after me.
• Built features that were shown in product demos and contributed to the Product Hunt launch.

Apr 2021 – Jul 2021 · 3 mos · Surat, Gujarat, India

• Joined in the founding month. Code I wrote in the internship shipped to production and was used by the platform's first users.