Skip to content
View shubham0086's full-sized avatar
🏠
Working from home
🏠
Working from home

Block or report shubham0086

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
shubham0086/README.md

Shubham Prajapati

AI Systems Builder · Forward Deployed Engineer · India

This repository hosts my personal portfolio. The best way to experience it is live, not as source files.

LinkedIn Email Open to Work


The Story

For 5 years I worked enterprise IT operations: SAP and ServiceNow, strict SLAs, the receiving end of broken software at HCL Technologies and EvoraIT. I learned what production discipline actually means.

Then, about 2 years ago, I started experimenting with AI on my own. No bootcamp, no team, no roadmap. I began by building small AI apps, taught myself as I went, and kept pushing until the apps became full systems.

The further I went, the harder the real problems got. Not "how do I call an LLM" problems, but systems problems:

  • Memory. My agents forgot everything between sessions. There was no continuity, no way to reuse what they had already learned.
  • Context. Agents wasted thousands of tokens re-reading entire codebases. Context windows blew up and so did the cost.
  • Reliability. A single-provider setup died the moment one API rate-limited or went down.
  • Consistency. Multi-agent pipelines drifted, contradicted themselves, and repeated the same failures in loops.

I solved each one inside my own codebases, then optimized: multi-provider routing with failover, persistent reality-first memory, context graphs to cut token waste, response caching, and guardrails against prompt injection and repeat failures.

Once those patterns were stable, I extracted them out of my private systems and open-sourced them, so other self-taught builders do not have to learn it all the hard way. This site is the showcase of that work.


The Journey

2019 - 2022   HCL Technologies, Noida
              L1 Agent to SME to Team Lead to Acting Floor Manager
              250+ agents trained, new support vertical live in 6 months, 4 promotions

2022 - 2024   EvoraIT Solution Pvt Ltd, Gurugram
              SAP L2 Support Specialist
              350 tickets/week, 98% SLA compliance, 95% MTTR improvement

2024 - Now    Self-taught AI engineering
              From first AI experiments to production systems and open-source patterns

I am open to Forward Deployed Engineer, Technical Solutions Engineer, and Founding Engineer roles at API-first AI startups.


What I Built

19 repositories · 261+ automated tests · 3 shipped products · 5+ years ops & SLA ownership

Full case studies, architecture diagrams, and live demos on the portfolio site — which itself runs a live RAG chatbot grounded on these repos (ask it anything).

Products

Product Stack Status
WellnessInYou + BODH — Full-stack AI wellness platform Next.js 15, Expo 54, CrewAI, Gemini, Razorpay/Stripe Live
Agentic SDLC — 6-agent software development pipeline Node.js, SQLite Blackboard, multi-LLM router, React+Vite Private beta
Agentic OS — Multi-agent outreach + content system FastAPI, Redis, LangChain, Instagram automation Internal
Adiyogi Books — Independent publishing brand Next.js, Prisma, Razorpay Live

Open-Source: Learn Track

Foundations for anyone building agent systems from scratch.

Repo What it teaches
AI-Systems-Evolution The same task at 6 levels of autonomy — code → single call → workflow → agent → team → swarm
Agent-Anatomy One agent, four removable organs (brain/hands/memory/loop). Toggle one off, watch it break

Open-Source: Solve Track

Production patterns extracted from real systems. Each repo isolates one hard problem.

Repo Problem solved Tests
Agent-Scars Failure memory — agents that learn from past mistakes 7
Agent-Context Dependency graphs to cut token waste on large codebases 6
Agent-Recall Cross-session solution memory with similarity retrieval 9
Agent-Routing Multi-provider LLM failover (NVIDIA → Groq → OpenRouter → Gemini) 23
Research-Agent Web research + multi-source summarization pipeline 9
Content-Analyzer URL → structured JSON extraction with AI analysis 4
Video-Engine-Starter Text-to-video orchestration with Remotion + AI directors
Agent-Constitution Drift detection + anti-hallucination guardrails 6
MCP-Agent-Toolkit MCP server with blackboard, SCAR, and cache tools (node:sqlite) 13
RAG-Knowledge-Engine Hybrid retrieval (BM25 + vector RRF) + cross-encoder rerank + RAGAS eval 25

Open-Source: Integrate Track

Complete systems and production templates.

Repo What it is Tests
Agentic-Patterns 7 architectural guides with runnable starters
Agentic-Systems 5 complete standalone agent systems
AgentKernel 6 production engines in Python + JavaScript
Agentic-SaaS-Boilerplate Billable multi-agent SaaS template — real WorkflowEngine, SSE queues, AgentRegistry 22

The Engineering Problems, and How I Solved Them

Problem What I built to solve it
Agents forgetting across sessions Reality-first persistent memory with cross-session reuse
Token waste from re-reading code Dependency context graphs and layered RAG filtering
Single point of failure on LLM calls Multi-provider router with circuit breakers and ordered failover
Brittle linear agent chains DAG orchestration with Kahn's topological scheduling
Agents drifting and repeating failures Anti-drift reality files and a repeat-failure guard (SCAR)
Prompt injection and secret leakage Input and output guardrails, tested against real payloads
RAG returning irrelevant chunks Hybrid BM25 + vector retrieval with RRF fusion + cross-encoder rerank
LLM provider downtime 5-provider failover chain — any key missing is silently skipped

Built With

HTML5 CSS3 JavaScript GitHub Pages

Static site. No framework, no build step. Hand-written HTML, CSS, and vanilla JavaScript with inline animated SVG diagrams. Loads in under a second. The portfolio chatbot runs as a Vercel serverless function — Gemini embeddings over a prebuilt 118-chunk index, generation on a 5-provider failover chain.


Contact

Pinned Loading

  1. agentkernel agentkernel Public

    Six production AI engines for multi-agent systems: Router, Memory, Retriever, Queue, Media, Auth. The infrastructure layer powering Agentic SDLC and Agency OS.

    Python

  2. agentic-systems agentic-systems Public

    Five complete, standalone agentic systems: customer support, research pipeline, code reviewer, data extractor, and content generator. Clone and run.

    JavaScript

  3. agentic-patterns agentic-patterns Public

    Seven production AI agent architecture patterns with runnable Node.js and Python starters: Reflection, Tool-Use, Planning, Multi-Agent, RAG, Memory, and Routing.

    Python 1

  4. wealthpath-app wealthpath-app Public

    Python

  5. Zapp-Builder Zapp-Builder Public

    Python