Skip to content
View 2004Snehal's full-sized avatar

Block or report 2004Snehal

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
2004Snehal/README.md

Hi, I'm Snehal Pal 👋

Generative AI engineer and researcher. I build RAG systems, LLM agents, Text-to-SQL agents and real-time voice AI.

B.Tech. (Hons.) Chemical Engineering, Micro Specialization in AI & Applications, IIT Kharagpur (2023–2027, CGPA 8.25/10).

Website LinkedIn Email arXiv


💼 Experience

  • Generative AI & Data Science Intern, Founders Office, Purplle.com (May–Jul 2026): Verification-first Text-to-SQL agent (LangGraph) automating 40+ financial P&L metrics; pgvector retrieval narrowing 120+ tables to 6 per query; SQLGlot validator with 96%+ pass rate against Metabase.
  • AI Engineer Intern, CoralComp (Mar–May 2026): AWS infrastructure, AI chatbot feature, faster data updates via star schema, Parquet on S3 and DynamoDB.
  • Co-Founder, Hodos (Mar–Jul 2025): B2B SaaS for corporate travel auditing. Multi-tenant RAPTOR retrieval engine and 4-hop reasoning pipeline with 95%+ verdict accuracy. Selected for AWS Activate and Microsoft for Startups.
  • Computer Vision Research Intern, IIT Kharagpur (Jan–Mar 2025): Complex-valued CNNs and multiscale complex ViT for PolSAR land-use classification.
  • NLP Research Intern, Mehdie.org, Aalborg University (Dec 2024–Jan 2025): Record linkage over 10M+ historical records with XLM-RoBERTa (LoRA).

📄 Research

Transaction Foundation Models for RAG over Financial Transaction Data (arXiv cs.IR, sole author) 1,026-question RAG benchmark over synthetic AML transactions. NVIDIA's 29M-param Transaction Foundation Model reaches 0.81 Recall@5 vs 0.25–0.39 for size-matched text embedders.

🚀 Featured Projects

Project What it does
voicedecide · PyPI Training-free call routing and spoken language ID on EmbeddingGemma 2. Zero/few-shot, returns act / confirm / escalate.
fn-dictate Hold fn to dictate anywhere on macOS. 100% local speech-to-text with IBM Granite Speech via Ollama.
Hodos RAG Microservice RAPTOR + multi-hop RAG for travel-policy audit. Context recall 0.61 → 0.91, faithfulness 0.93 (RAGAS).
RAG Optimization Latency optimization experiments for RAG pipelines.
Speaking Meeting Bot Autonomous speaking bots for meetings with Pipecat.
dsa-tracker LeetCode tracker: progress by topic, pattern notes, C++ templates. Next.js on Vercel.

🛠️ Tech Stack

AI/ML: PyTorch · LangGraph · LangChain · Hugging Face · RAG (RAPTOR, pgvector) · LoRA · Pipecat · Deepgram · Groq Languages: Python · TypeScript · C++ · SQL · C# Infra: AWS (ECS, S3, DynamoDB) · PostgreSQL · Docker · Vercel · Next.js · React

🏆 Achievements

  • Selected for Y Combinator Startup School (out of 20,000 applicants)
  • Top 1% in JEE Advanced 2023
  • INMO Camp 2021 · Codeforces Specialist (1400+)
  • Bronze, General Championship Data Analytics · Open IIT Supply Chain Bronze (twice)

📫 Open to GenAI / ML engineering roles and research collaborations. Reach me at palsnehal004@gmail.com.

Pinned Loading

  1. hodos_microservice_rag hodos_microservice_rag Public

    Python 1

  2. neetcode-submissions neetcode-submissions Public

    My NeetCode.io problem submissions

    C++

  3. pror993/Ekta pror993/Ekta Public

    TypeScript

  4. voicedecide voicedecide Public

    Training-free audio decisions (call routing, language ID) with EmbeddingGemma 2

    Python 6