AI Engineering
01Model-backed features that hold up outside the demo.
- RAG
- Evals
- Prompt Engineering
- Vector Search
- OpenAI
hey, i’m
AYUSH.
↻ give the panel a tap
I buildAI SYSTEMSAGENTSWEB APPSDEV TOOLSpeople actually use
RAGEvalsPrompt EngineeringVector SearchOpenAILangGraphMCPTool CallingMulti-Agent OrchestrationNode.jsTypeScriptNext.jsReactTailwind CSSPostgreSQLDrizzleMongoDBPrismaGroqHugging FaceProvider Fallback ChainsSSE Streaming
01About
Engineering the part where a model becomes a product.
I am an engineer who cares about the seam between models and products — the part where a demo becomes something people rely on every day.
Most of my work sits between AI systems and the interfaces that make them legible: retrieval pipelines, agent graphs, evaluation harnesses, and the front ends that keep all of it honest.
When I am not shipping, I am usually reading distributed systems papers or grinding LeetCode with a coffee that went cold two hours ago.
- Public repositories
- 64
- Problems solved
- 379
- Submissions this year
- 828
- GitHub followers
- 27
Ayush Kumar
AI Engineer
Capabilities
What I actually build with
No ratings out of ten. Every group below is something the work in this page runs on.
Agentic AI
02Stateful agents with explicit handoffs and traceable runs.
- LangGraph
- MCP
- Tool Calling
- Multi-Agent Orchestration
Backend
03APIs designed to survive their second year.
- Node.js
- TypeScript
- Next.js
Frontend
04Interfaces with motion that means something.
- React
- Next.js
- Tailwind CSS
Database & ORM
05Modelling data for the queries it will actually get.
- PostgreSQL
- Drizzle
- MongoDB
- Prisma
Models & Serving
06Running models in production, where the provider is the flaky part.
- Groq
- Hugging Face
- Provider Fallback Chains
- SSE Streaming
- Model Routing
Experience
Where I have been responsible for the thing working
- 01
SMVIT Debating Society
smvitdebsoc.comTook the society from a promotional page to the application the cabinet actually runs sessions on — public site, members' area and the AI practice features layered on the session data.
- Role-separated access: members see their own record, the cabinet sees the society
- Session conduct flow used live during club meetings
- Prisma over PostgreSQL for member, session and attendance records
- 02
Wedsy
wedsy.inA live commercial product I maintained on both sides of the wire. Straight web engineering rather than an AI build — most of the weight sat in the backend and the API surface the front end runs on.
- Maintained the frontend and the backend behind it, not one half of the stack
- Complex backend: a wide API surface driving every screen on the site
- Shipping against a client's production site, not a sandbox
Flagship
I built the assistant that coaches me from evidence, instead of asking how confident I feel.
An open-source mentor backend that turns your real work into interview-readiness coaching — watching Telegram reflections, GitHub activity and LeetCode practice, then coaching you on a schedule instead of behaving like a stateless chatbot.
- Morning
Checks in
Opens the day over Telegram and asks what the plan is, before there is anything to rationalise.
- All day
Collects the evidence
Telegram reflections, GitHub commits and LeetCode submissions land as activity events. Nothing is self-reported.
- Evening
Reflects against what happened
The evening workflow reads the day's events back, not my summary of them.
- Weekly
Scores and re-plans
Readiness is recomputed in code from the week's evidence, and the next week's plan is written against that number.
Portfolio
Selected work
Five systems, each one still running. Open any of them for the decisions behind it.
SMVIT DebSoc
The website for my college's debating club — a rich front end backed by a real system that took it from a promotional page to the tool the club actually runs sessions on, with AI features that help members upskill their debating.
- sessions
- conduct flow run live during club meetings
- members
- roster, participation and history in one place
- access
- role-separated — members see their own record, the cabinet sees the society
- stack
- Next.js App Router, Prisma/Postgres, Tailwind
- 100+
- concurrent users
- 0
- failed requests
- role-gated
- cabinet vs member
PathwayAI
An open-source mentor backend that turns your real work into interview-readiness coaching — watching Telegram reflections, GitHub activity and LeetCode practice, then coaching you on a schedule instead of behaving like a stateless chatbot.
- workflows
- morning check-in, evening reflection, weekly review
- evidence
- Telegram reflections, GitHub commits, LeetCode submissions
- scoring
- deterministic, computed in code rather than by the model
- stack
- FastAPI, LangGraph, PostgreSQL, Groq
- 3
- scheduled workflows
- deterministic
- readiness score
- groq + hf
- model lanes
RepoPilot
Paste a public GitHub URL, say why you are here, and get a tour of the codebase shaped by that intent — with every factual claim tied to a source span and re-checked by a verifier before you see it.
- intent
- persona or free text, captured before any analysis runs
- retrieval
- dense pgvector plus BM25, cross-encoder rerank, MMR
- claims
- every one cites a line range and is verified before display
- stack
- FastAPI, Next.js, LangGraph, pgvector, tree-sitter
- file:line
- every claim
- verified
- before display
- SSE
- streamed while indexing
CodeBuddy
Durable memory and prepared context for Claude Code and Codex: knowledge lives as readable Markdown in your repo, and the agent gets the smallest useful slice of it before every edit instead of re-reading the project each turn.
- memory
- Markdown under .codebuddy/, committable and reviewable
- surfaces
- CLI, MCP stdio server, TypeScript SDK, LangGraph helpers
- index
- PostgreSQL + pgvector, rebuildable from the Markdown
- stack
- TypeScript, Node.js, MCP, pgvector
- npm
- published
- local-first
- nothing leaves the machine
- 4
- surfaces, one core
ClawCode
A Telegram-controlled autonomous coding agent: one trusted operator sends a task, the agent inspects the repo, edits code and runs tests in an E2B sandbox, and opens a pull request only after explicit approval.
- control
- Telegram commands, free text and voice notes
- sandbox
- E2B for isolated dependency install and pytest runs
- gate
- human approval required before any push or PR
- stack
- Python, LangGraph, E2B, GitHub API
- phone
- approval gate
- checkpointed
- every dispatched tool
- sandboxed
- tests before proposal
Writing@ayushkumar320
Writing
My
Writing
Swipe → · Tap a card
05Signals
The public record.
Problem solving and open source, pulled live where the APIs allow it.
LeetCode
Live379
Problems solved
- Rating
- 1651
- Rank
- 346,073
156
Easy
198
Medium
25
Hard
Submissions828 in the last year
LessMore
GitHub
LiveAyush Kumar
@ayushkumar320Just a beginner
Bengaluru, Karnataka, IndiaJoined Dec 202027 followers38 following64 repositories
06Contact
Let’s build something worth shipping.
Open to AI engineering and full stack roles, freelance builds and genuinely interesting problems. I reply to everything.