Skip to content
View SujitSpark's full-sized avatar

Highlights

  • Pro

Block or report SujitSpark

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.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. 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
Sujitspark/README.md

hey, I'm Sujit

Terminal: whoami prints Sujit Shiravle, AI and full-stack developer. show-evals knowledgebot prints citation accuracy 15/15 (was 10/15), off-topic refusal 5/5 (was 4/5), red-team 10/10 (was 6/7). 3 bugs caught, 3 fixed, 3 regression tests.

Final-year B.Tech AI & Data Science · Navi Mumbai · I build LLM apps, then I try to break them.


🧪 How I build AI

Pipeline: question, retrieve with pgvector, generate with an LLM, guardrails, judge with DeepEval, cited answer.

🔬 My 5-step playbook for testing an LLM app (click to open)
  1. Write the exam first. A golden dataset with real answers, questions the bot must refuse, and attacks.
  2. Test the real thing. Evals hit the live API, not a copy of the pipeline.
  3. Trust, but calibrate. Cheap deterministic checks gate a release; LLM-as-judge scores are a signal I check against human review.
  4. Attack with phrasings I didn't train on. If the guardrail only stops the attacks I wrote, it isn't a guardrail.
  5. Every bug becomes a regression test, and CI runs them on every push.

That's how KnowledgeBot went from 67% to 100% citation accuracy, and how a benchmark caught JudicAI's risk model reading its own answer key.


🚀 Things I've built

🛡️ KnowledgeBot

Chat with your PDFs, 100% offline. React Native PostGraphile pgvector Ollama

🧪 DeepEval + promptfoo eval and red-team suite · guardrails · CI

⚖️ JudicAI

Multi-agent triage for court backlogs. LangGraph FastAPI React XGBoost Gemini

📊 1,200-case benchmark · leakage bug found and fixed · AUC 0.60 → 0.65

Game theory for food-delivery price wars: Nash, Bertrand, Bayesian and repeated games. FastAPI React TypeScript

Detects 27 fine-grained emotions, picks an empathy strategy, then lets an LLM reply. Team project. PyTorch Transformers Gradio


📜 git log --oneline --career

+ 2026-11  feat(evals): red-team suite for KnowledgeBot, 3 bugs caught, 3 fixed
+ 2026-07  docs(research): co-authored IEEE I2ITCON 2026 paper on an AutoML learning platform
+ 2026-06  feat: KnowledgeBot, an offline RAG assistant (pgvector + Ollama)
+ 2026-03  ctf: 4th place, Technovate'26 Capture The Flag
+ 2026-02  feat: JudicAI, multi-agent judicial triage (LangGraph + FastAPI)
@ 2025-10  job: SDE Intern @ Infidhi Ventures (React Native + GraphQL/PostgreSQL)
@ 2025-08  job: Junior Intern @ Infidhi Ventures (PERN stack + RAG over internal docs)
@ 2024-12  job: Web Developer Intern @ Nexon Logistics (Firebase event booking)
+ 2024-06  feat: SIES GST Student Portal, used by 4,000+ students, 12,000+ forms
+ 2024     win: 2nd Runner Up, SustainX 2024
  2023-06  init: B.Tech AI & Data Science @ SIES GST

🧰 Toolbox

Python, TypeScript, JavaScript, Java, C++, React, Node.js, Express, FastAPI, GraphQL, PostgreSQL, MongoDB, Firebase, Docker, Kubernetes, Git, GitHub


Contribution snake

⚡ Ask me about: why an R² of 1.00 is a red flag · making a chatbot say "PWNED" · RAG that runs with no internet

Popular repositories Loading

  1. JudicAi JudicAi Public

    AI-powered judicial intelligence system with multi-agent orchestration for case prioritization, risk scoring, and automated memo generation

    TypeScript 1

  2. knowledgebot knowledgebot Public

    Offline RAG-powered knowledge assistant using React Native, PostGraphile, pgvector, and Ollama — no cloud APIs required.

    TypeScript

  3. NLP NLP Public

    Python

  4. Enigma_BrewtForce Enigma_BrewtForce Public

    JavaScript

  5. price_war price_war Public

    TypeScript

  6. Sujitspark Sujitspark Public