Skip to content
View KshitijAng's full-sized avatar

Block or report KshitijAng

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
KshitijAng/README.md

Hi, I'm Kshitij 👋

Early-career engineer interested in building LLM-backed systems with backend depth - orchestration, retrieval, observability, and thinking about where AI should and shouldn't make decisions.

Open to Applied AI / AI Engineer roles

Experience

Hiver - AI Intern · Dec 2025 – Jun 2026

  • Assisted in developing a 5-phase LLM pipeline to surface AI adoption opportunities across the customer base, using batched GPT-5.4 calls with retry backoff, signal-keyword filtering, and PII-aware email preprocessing.

  • Supported development of an SQS-based orchestration pipeline computing per-usergroup AI usage metrics across 30/60/90-day windows, publishing feature-wise JSON snapshots to S3 with Redis-precomputed aggregations.

  • Contributed to an end-to-end AI automation testing platform (React + FastAPI + Playwright), enabling automated execution of a 15-step Chat Agent scenario with per-run artifact capture (replay video + failure screenshots).

Ernst & Young - Generative AI Intern · Sep 2024 – Nov 2024

  • Built a document Q&A RAG pipeline using LangChain, Stella embeddings selected via MTEB leaderboard, and Chroma, with PyMuPDF text extraction, chunking, and top-k retrieval for grounded source-cited answers.

  • Benchmarked local Llama 3.1 against Groq-hosted inference on the same pipeline, evaluating the privacy vs. latency tradeoff for sensitive document workloads.

Kreativstorm - Data Analysis Trainee · Sep 2024 – Oct 2024

  • Cleaned and analysed 8+ datasets using IBM SPSS, resolving inconsistencies to improve data reliability and produce decision-ready outputs.

  • Collaborated with a cross-functional remote team to streamline reporting workflows and deliver data-driven project insights.

What I've built

🤖 Dossier - Agentic AI research assistant built on a 5-node LangGraph state machine (planner → searcher → reader → synthesizer → critic) with a bounded self-critique loop, citation-grounded reports, Langfuse observability, and a Next.js + MUI frontend. Live ↗

🔍 Hybrid Search System - Hybrid retrieval over 5,753 chunks using dense + sparse search combined via rank-based fusion, Cohere rerank-3.5 cross-encoder, citation-grounded generation, and end-to-end Langfuse tracing. Article ↗

Vendor Onboarding Validator - 6-stage deterministic + LLM hybrid validation pipeline with hard FAIL vs soft WARN logic. LLM restricted to borderline name-mismatch adjudication only - deterministic checks handle all upstream validation.

🎫 TicketSense - Groq-powered LLM ticket triage processing 300 tickets in ~12 minutes, backed by layered async Python architecture and two-tier Postgres + Redis storage.

Stack I work with

  • Languages: Python · SQL
  • AI / LLM: LangChain · LangGraph · Langfuse · RAG · Pinecone · pgvector · Groq
  • Backend: FastAPI · Pydantic · SQLAlchemy · SQS · AWS (S3, Athena)
  • Databases: PostgreSQL · Redis · MySQL · Elasticsearch
  • Infra: Docker · Git

What I'm focused on

  • Agentic AI - LangGraph state machines, bounded self-critique loops, tool-use orchestration
  • Hybrid retrieval - dense + sparse fusion (RRF) and cross-encoder reranking
  • Deterministic-first AI design - LLMs for ambiguity, deterministic logic for everything else
  • LLM observability - Langfuse tracing across every node, token cost and latency per request
  • Citation-grounded generation - making LLM answers traceable to source chunks

Connect

LinkedIn · kshitijangurala903@gmail.com · Linktree

Pinned Loading

  1. Hybrid-Search-System-pgvector-Cohere-Rerank-3.5 Hybrid-Search-System-pgvector-Cohere-Rerank-3.5 Public

    Hybrid Search RAG over technical docs - pgvector + Postgres FTS fused via Reciprocal Rank Fusion, Cohere rerank-3.5, citation-grounded answers, Langfuse traces.

    Python

  2. TicketSense TicketSense Public

    LLM-powered support ticket triage. Groq with primary + fallback models, Pydantic structured outputs, Postgres + Redis two-tier storage, Streamlit dashboard.

    Python

  3. FlowVentry-Inventory-Management FlowVentry-Inventory-Management Public

    Next.js 15 + MongoDB inventory app - module-level cached MongoClient singleton, $text aggregation pipeline with a compound text index for relevance search.

    JavaScript

  4. AI-Hub-SpringBoot AI-Hub-SpringBoot Public

    Spring Boot 3.5 + React 19 chat + code-generation app on Spring AI and Groq's OpenAI-compatible API, with reusable PromptTemplate definitions across routes.

    Java

  5. Churn-Prediction-Decision-Tree Churn-Prediction-Decision-Tree Public

    Class-balanced decision-tree churn classifier (one-hot, CV-tuned max_depth) - lifted churn-class recall 0.43 → 0.77 over a 4-feature baseline. FastAPI + ONNX.

    Jupyter Notebook