Darshit Sengra

Software Engineer · Forward Deployed Engineer · AI / full-stack

I build autonomous agents, ML pipelines, and full-stack products — from a causal SRE platform to real-time voice agents. This realm is my portfolio: every building below enshrines a project you can inspect.

Currently building

Apiary

An autonomous coding-agent fleet that fixes bounded issues and opens PRs — never merging on its own.

My main build right now. A deterministic control plane (bash + systemd + SQLite) gates which issues are safe and dispatches ephemeral Claude agents to fix the bounded ones in isolated worktrees — the LLM only writes the fix, so safety is code, not prompts. Multi-tenant SaaS with structurally-enforced org isolation. Honest stage: the single-tenant loop is proven end-to-end on a live EC2 server; per-tenant execution and billing are what's left before paying clients.

Finddder

An internship & opportunity tracker that scrapes, ranks, and surfaces roles.

An internship and opportunity tracker that scrapes listings, ranks them, and surfaces the best matches behind a typed API and a fast React front end. Actively building.

Projects

24 shipped and in-flight projects across AI and agents, systems and infrastructure, data products, and full-stack ventures.

KubeStellar + Kyverno

open source

CNCF / LFX contributor with a real merge record.

Open-source contributions across two CNCF projects: 10+ merged PRs and 25+ filed issues spanning the KubeStellar Console and Kyverno. The work came from deep manual navigation of unfamiliar codebases and was mentor-endorsed through the LFX program.

  • React
  • TypeScript
  • Go
  • CNCF
  • Kubernetes

JEE Arena

venture

A LeetCode-style practice arena for JEE aspirants.

A competitive practice arena for India's JEE exam: timed problem sets, submissions, and a judging pipeline that scores attempts. An ongoing venture built with a co-founder. In Devhalla, this project literally is the combat arena.

  • Node.js
  • TypeScript
  • FastAPI
  • Python

Voice Agent (revRAG.ai)

production

Production real-time voice-agent pipelines.

Production real-time voice agents: speech and telephony triggering, plus a WebSocket account-aggregator integration handling consent and OTP flows. Built for revRAG.ai. In the realm, the Oracle speaks this blurb aloud.

  • LiveKit
  • WebSockets
  • Python
  • Telephony

Nemotron RL

research

Full fine-tuning + RL pipeline on a Nemotron-3 Nano model.

An end-to-end fine-tuning and reinforcement-learning pipeline on a Nemotron-3 Nano model: supervised fine-tuning, GRPO, LoRA adapters, and vLLM inference, with a symbolic-solver fallback for hard math reasoning. The familiar in the Spire 'levels up' along the reward curve.

  • GRPO
  • SFT
  • LoRA
  • vLLM
  • Mamba

Finddder

product

An internship tracker that scrapes, ranks and surfaces roles.

An internship and opportunity tracker that scrapes listings, ranks them, and surfaces the best matches behind a typed API and a fast React front end. The market merchant 'scouts' roles the way the product does.

  • React
  • TypeScript
  • FastAPI
  • Python

Multi-Agent RAG

research

A multi-agent FastAPI RAG system over a document corpus.

A FastAPI service that answers questions over an ingested document corpus (built against Hyperledger Fabric docs) using Google Generative AI for embeddings and generation, with ChromaDB as the vector store. Separate ingestion and chat workflows handle document management and retrieval-augmented query answering.

  • FastAPI
  • Google Generative AI
  • ChromaDB
  • Python
  • RAG

OpenEnv — Causal SRE Platform

production

RL simulation platform for autonomous SRE incident response.

A production-grade reinforcement-learning simulation platform for autonomous SRE incident response, modeling microservice architectures with realistic anomaly injection. A hybrid ML pipeline — XGBoost (485-feature root-cause classification), Isolation Forest, and an LSTM autoencoder — feeds a causal reasoning engine (PC-algorithm DAG discovery + an SEM counterfactual engine) for safe 'what-if' intervention, with SHAP explainability.

  • FastAPI
  • React
  • Docker
  • XGBoost
  • PyTorch
  • SHAP

BTC Price-Range Predictor

research

95% next-hour BTC price-range forecasts via Monte Carlo.

A GBM model forecasting the 95% BTC/USDT confidence interval via Monte Carlo (10k paths, Student-t log-returns) — 95.28% coverage on a 30-day backtest. It benchmarks three volatility models (Simple σ, EWM, HAR-RV), tunes hyperparameters by Winkler score, and ships a live Binance feed with SQLite persistence behind an interactive Streamlit + Plotly dashboard.

  • Python
  • Streamlit
  • Plotly
  • Monte Carlo
  • SQLite
  • Binance API

infLLM — Short-Form Video RAG

product

Ask questions across YouTube & Instagram videos with cited answers.

Paste YouTube and Instagram URLs and the app pulls their transcripts, embeds them with Gemini, and stores the vectors in ChromaDB for cosine retrieval. A LangGraph flow does a dual-retrieve pass and answers with gemini-2.5-flash, streaming the response over SSE into a Next.js UI with timestamped citation chips back to the source video.

  • Next.js
  • LangGraph
  • Gemini
  • ChromaDB
  • SSE
  • Python

Resume Enhancer

product

A multi-agent résumé reviewer with pluggable LLM providers.

A Streamlit app that reviews a résumé through a multi-agent pipeline built on LangChain and a langgraph StateGraph. It reads the PDF with pypdf and is provider-agnostic — Gemini, OpenAI, OpenRouter, or Groq — so the same review graph can run on whichever model you have keys for.

  • LangChain
  • langgraph
  • Streamlit
  • pypdf
  • Python

Smart Study Assistant

product

An AI study companion: summaries, MCQs, and a math mode.

A study helper that pulls a Wikipedia summary for a topic and has gemini-2.5-flash generate a 3-bullet summary, three multiple-choice questions, and a study tip. It includes a math mode, a history of past lookups, and a dark/light theme, served by a small Node/Express backend.

  • Node.js
  • Express
  • Gemini
  • Wikipedia API
  • JavaScript

Project Echo

product

A Flask RAG chatbot over a source PDF.

A retrieval-augmented chatbot that loads a source PDF with PyPDFLoader, embeds it with Azure OpenAI, and stores the chunks in ChromaDB. A Flask web UI lets you chat against the document, retrieving relevant passages before composing each answer.

  • Flask
  • Azure OpenAI
  • ChromaDB
  • PyPDFLoader
  • Python

Hand-Gesture Control

research

An early computer-vision project: control the cursor by hand.

A first computer-vision experiment that tracks hand landmarks with OpenCV and MediaPipe and maps gestures to mouse and keyboard actions via pyautogui. Small and exploratory — the project where the CV fundamentals were learned.

  • OpenCV
  • MediaPipe
  • pyautogui
  • Python

Women Safety App

venture

Hackathon-winning women's safety app (Polaris & Google for Developers).

The hackathon-winning women's safety application — 1st place at the Polaris & Google for Developers hackathon. A React + TypeScript front end backed by a Node server, with an SOS button, fake-call decoy, live location tracking, a geofencing layer, emergency contacts, community alerts, an AI chat assistant, a map view, and auth.

  • React
  • TypeScript
  • Node.js
  • Express

Mock E-Commerce Cart

venture

A MERN shopping cart over the Fake Store API.

A mock e-commerce cart built on the MERN stack — a React/Vite front end with an Express + MongoDB backend, pulling its catalog from the Fake Store API. A full-stack practice build covering product listing and cart state.

  • React
  • Vite
  • Express
  • MongoDB
  • Node.js

Mini Twitter

venture

A small MERN microblog with a posts/notes API.

A minimal Twitter-style microblog on the MERN stack: an Express notes/posts API with React components for creating and listing posts. A compact full-stack exercise in CRUD and component state.

  • React
  • Express
  • MongoDB
  • Node.js

Horse-Betting Game

venture

A simple browser horse-betting game.

A small JavaScript browser game: bet on one of five horses that race at random speeds to a finish line. A quick, self-contained front-end exercise.

  • JavaScript
  • HTML
  • CSS

Finance Dashboard

product

A TypeScript finance dashboard with a split front/back end.

A finance dashboard built in TypeScript with a separate frontend and backend. A practice full-stack data app focused on presenting financial figures in a dashboard layout.

  • TypeScript
  • Node.js
  • React

FastAPI Items CRUD

product

A small items CRUD REST API, deployed live on Vercel.

A compact CRUD REST API for items built with FastAPI and deployed live on Vercel. A clean reference for a typed Python API with create/read/update/delete endpoints.

  • FastAPI
  • Python
  • Vercel

Forest-Fire Predictor

research

Ridge regression on the Algerian forest-fires dataset, served by Flask.

A Flask app that predicts a forest-fire indicator using a Ridge regression model trained on the Algerian forest fires dataset. The trained model and its scaler are persisted as pickles and loaded to serve predictions from a simple web form.

  • Flask
  • scikit-learn
  • Ridge regression
  • Python

California Housing Regression

research

A learning exercise: multiple linear regression on California housing.

A small learning project applying multiple linear regression to the California housing dataset — fitting and evaluating a baseline model to practice the regression workflow end to end.

  • scikit-learn
  • pandas
  • NumPy
  • Python

CRM REST API (Go Fiber)

production

A CRM REST API built with Go, Fiber, GORM, and SQLite.

A CRM REST API in Go using the Fiber web framework with GORM over SQLite for persistence. A clean back-end build covering the standard CRUD contact/lead endpoints.

  • Go
  • Fiber
  • GORM
  • SQLite

Liquid Galaxy — GSoC Task 2

open source

A Flutter qualification task for Liquid Galaxy (GSoC 2025).

A Flutter app submitted as a GSoC 2025 qualification task for the Liquid Galaxy org. Part of the application process for contributing to the multi-screen Liquid Galaxy rig software.

  • Flutter
  • Dart

Liquid Galaxy — GSoC Task 5

open source

A second Flutter qualification task for Liquid Galaxy (GSoC 2025).

A second Flutter qualification task for the Liquid Galaxy org's GSoC 2025 application, building on the org's task track for prospective contributors.

  • Flutter
  • Dart

Skills & technologies

  • Python
  • AI / ML
  • Reinforcement Learning
  • LLM Engineering
  • GRPO
  • LoRA
  • vLLM
  • SFT
  • TypeScript
  • Frontend
  • React
  • Three.js / R3F
  • Go
  • Cloud / CNCF
  • Kubernetes
  • KubeStellar
  • Kyverno
  • Node.js
  • Backend / APIs
  • FastAPI
  • Realtime
  • LiveKit
  • WebSockets

Honours & awards

  • Hackathon Winner — Polaris & Google for Developers — Feb 20251st place for an innovative women's safety application.
  • WorldQuant International Quant Championship — 2026Top 20% of teams globally in Stage 1 (team "Cornflakes").
  • AWS Cloud Practitioner EssentialsCore AWS services, security, architecture, and pricing.
  • Competitive ProgrammingSolved 450+ algorithmic problems on LeetCode.

The path so far

  • Cutting my teeth — The grindBefore any title, there were the reps — 450+ algorithmic problems on LeetCode. That's where I learned hard problems yield to patience and pattern, not cleverness.
  • First merges in the open — Into the openI earned my way into CNCF work through the LFX program: 10+ merged pull requests and 25+ issues across KubeStellar and Kyverno, mentor-endorsed. I learned to read a codebase I didn't write and ship into it — in public.
  • Voices in the wire — revRAG — ProductionAt revRAG.ai I built production real-time voice agents: speech and telephony triggering, plus consent and OTP flows over WebSockets. Real users, real uptime — the kind of work that has to answer the phone at 3am.
  • First place, real stakes — Feb 2025A women's safety application took 1st at the Polaris & Google for Developers hackathon — building something that actually matters, fast, with a team.
  • Against the world's quants — 2026Team "Cornflakes" finished in the top 20% globally in Stage 1 of the WorldQuant International Quant Championship — proof the pattern-work pays off under pressure.

Questions people ask

Who is Darshit Sengra?
Darshit Sengra is a software engineer working across AI/ML and full-stack engineering. I build autonomous agents, ML pipelines, and full-stack products — from a causal SRE platform to real-time voice agents.
What is Devhalla?
Devhalla is the software engineer portfolio of Darshit Sengra, built as a playable 3D RPG world. Instead of a scrolling list of projects, every project is a building you can walk into and inspect, and the characters answer questions about the work in character. It runs in the browser on Three.js and React with a FastAPI backend.
What is Darshit Sengra working on now?
Apiary — An autonomous coding-agent fleet that fixes bounded issues and opens PRs — never merging on its own. Finddder — An internship & opportunity tracker that scrapes, ranks, and surfaces roles.
What technologies does Darshit Sengra use?
Python, AI / ML, Reinforcement Learning, LLM Engineering, GRPO, LoRA, vLLM, SFT, TypeScript, Frontend, React, Three.js / R3F, Go, Cloud / CNCF, Kubernetes, KubeStellar, Kyverno, Node.js, Backend / APIs, FastAPI, Realtime, LiveKit, WebSockets.
How do I contact Darshit Sengra?
By email at darshit02736@gmail.com, or through GitHub at https://github.com/Darshit42. A downloadable resume (PDF) is at https://www.darshit.wtf/resume.pdf.

Away from the keyboard

Away from the keyboard: he plays guitar, sings, and plays piano; he writes poetry and sketches; and he cooks.