# Supril Singh
Supril Singh builds production AI agents, cloud-native platforms, backend systems, and full-stack developer tools.
## Navigation
- [Human site](/)
- [Blog](/blog/)
- [Machine interface](/machine/)
- [Raw Markdown site map](/site.md)
- [Sitemap](/sitemap-index.xml)
## Profile
- Role: Senior Software Engineer / Applied AI Engineer
- Location: Pune, Maharashtra, India
- Email: suprilsingh12@gmail.com
- Phone: +91 7972057832
- LinkedIn: https://linkedin.com/in/supril-singh
- Summary: I am trying to build the machine that develops other machines. Call it humanity's last piece of software: agents, runtimes, and platforms that can understand systems, use tools, and help ship the next layer of software.
- Availability: Open to applied AI, AI platform, backend, founding engineer, and forward-deployed engineering roles.
## Highlights
- Around 5 years building backend platforms, AI systems, cloud infrastructure, and full-stack products.
- Production experience with LangGraph, OpenAI tool calling, MCP-style architectures, Vercel AI SDK, and sandboxed TypeScript execution.
- Comfortable owning the whole path from messy system behavior to APIs, workers, dashboards, auth, audit trails, and launch.
- Strong fit for high-ownership roles where AI, platform engineering, and product execution meet.
## Built and Shipped
- RLM-style Codebase Analyzer (2026). I built an experimental analyzer for large, multi-file codebases where the model does not need the whole repo pasted into context. It inspects files through code, narrows the search space, runs focused passes, and keeps only the useful findings in the main reasoning loop. Proof points: Used the RLM idea of keeping huge inputs outside the model context and letting programmatic inspection pull in only the parts that matter. Split big-codebase investigation into smaller passes: locate relevant files, summarize local behavior, compare call paths, and then synthesize likely root causes. Designed it for multi-file debugging and architecture review, where the hard part is not reading one file but staying coherent across a large codebase. Reference: Recursive Language Models by Prime Intellect (https://www.primeintellect.ai/blog/rlm). Stack: Recursive Language Models, Python, LLM orchestration, code analysis, context folding.
- Traceminer (2026): https://traceminer.dev. I built Traceminer to crack open undocumented APIs from messy HAR files and turn traffic noise into clean maps, SDKs, and replayable workflows. Proof points: Handles large traces, reconstructs private API behavior, and gives developers something they can actually test instead of guessing from browser calls. Uses agentic workflows and sandboxed code execution so the system can reason through ugly traffic without turning every step into prompt soup. I shipped the product surface, backend orchestration, SDK generation, replay tooling, payments, and the developer workflow end to end. Stack: Cloudflare Workers, React, Astro, Hono, Vercel AI SDK, OpenRouter, TypeScript.
- Creator Automation Platform (2026). A Telegram-first AI studio that turns rough ideas, links, and reference clips into short-form videos without a manual production line. Proof points: Creators drop prompts, references, and source links in Telegram; the system ingests, queues, extracts media, and moves the job through the pipeline. Local and serverless AI workers generate scripts, voices, captions, subtitle timing, compositions, and final renders. Built the queueing, retries, publishing flow, and GPU-heavy inference boundaries so the pipeline can keep moving without babysitting. Stack: Cloudflare Workers, Telegram Bots, Python, TypeScript, Open Source LLMs, TTS, Voice Cloning, Video Rendering.
## Experience
- Senior Cloud Application Developer, ZS Associates. Jan 2024 - Present. Architected and developed Zeus, a production monitoring platform with FastAPI, React, and role-based access control. Developed MST, a MicroStrategy migration automation system that handles end-to-end migration workflows and reduces manual handoffs across the process. Building an agentic Smart Assist chatbot that can be embedded inside Power BI dashboards and answer business questions by talking directly to Snowflake. Designed workflow agents that query internal APIs, transform data, and execute multi-step operations with schema validation, permission checks, structured outputs, and audit logs. Launched a license reconciliation system that automates entitlement checks, reminders, tracking, and deadline-based access revocation.
- Cloud Application Developer, ZS Associates. Aug 2021 - Dec 2023. Built Java microservices for node-level health checks on AWS Lambda, Parameter Store, S3, and API Gateway, reaching 99.9% availability. Implemented a custom MicroStrategy SSO plugin with MSTR SDK, Open token, and SAML-based authentication. Designed the Chargeback module database schema and led 3 engineers through implementation with Spring Boot, Java, and Hibernate. Built a Python migration and transformation system across Power BI and Tableau that cut admin work by 90% and reduced turnaround from 6 hours to 1 hour.
## Expertise
- Applied AI systems: AI agents, agentic workflows, multi-agent systems, tool calling, structured outputs, LLM orchestration, MCP, LangGraph, OpenAI APIs, Vercel AI SDK, sandboxed execution.
- Backend and platform: Python, Java, TypeScript, FastAPI, Spring Boot, Hono, REST APIs, microservices, distributed systems, RBAC, JWT, SSO, audit logging.
- Cloud and product: AWS Lambda, API Gateway, S3, Parameter Store, Cloudflare Workers, KV, D1, R2, Durable Objects, Docker, CI/CD, React, Astro.
## Good Fit For
- Applied AI Engineer
- AI Platform Engineer
- LLM Engineer
- AI Runtime Engineer
- Agentic Systems Engineer
- Senior Backend Engineer
- Platform Engineer
- DevTools Engineer
- Full-Stack AI Engineer
- Founding Engineer
- Forward Deployed Engineer
- AI Solutions Engineer
## Education
- Bachelor of Engineering in Information Technology, RCOEM, Nagpur. May 2021. CGPA: 7.89.
## Blog Posts
- [How I Reverse Engineered LinkedIn's Bot Detection and Parameter X](/blog/parameter-x-linkedin-bot-detection/): I tore apart LinkedIn's obfuscated PerimeterX security script, 14,130 lines, 440+ encoded strings, and a fingerprinting engine that watches your every move, to understand what "Parameter X" really was. Category: Engineering. Published: Sep 20, 2025. Tags: bot-detection, reverse-engineering, automation, perimeterx, javascript.
## Agent Notes
- This site is static Astro with no frontend framework.
- The visual system is controlled from src/styles/global.css.
- Site data is controlled from src/config/site.ts.
- Blog content is Markdown in src/content/blog/.
- Prefer semantic HTML, small CSS, fast pages, and minimal dependencies.