Siddhant Kaushik

about me.

Hey, I’m Siddhant — I like turning rough ideas into working tools across backend, AI/ML, and smart-contract stacks. Lately focused on safety, automation, and trustworthy blockchain rails. Shipping fast with a bias for reliable systems: AI safety signals (EcoSentry), mempool anomaly checks (ChainSentinel), and resilient communication for no-network zones (Atithi Rakshak). Open to product-focused backend/ML roles and hackathons.

proof of work.

timeline.

• Dev Fellow Formstr-org · Remote

  at Bitshala

  JUN 2026 - PRESENT

  • Contributing to Formstr-drive and Formstr docs, building privacy-preserving decentralized file storage alternatives to traditional cloud platforms.
  • Working on Nostr and Blossom infrastructure for censorship-resistant, user-owned storage.

  • Nostr
  • Blossom
  • Decentralized Storage
  • Open Source

• GSoC Contributor, AOSSIE Remote

  at Google summer of code

  MAY 2026 - PRESENT

  • Engineering the core architecture of a decentralized blockchain protocol, integrating PoW consensus, libp2p secure P2P networking, and optimized mempool mechanics for scalable state persistence.
  • Implementing cryptographic transaction signing utilizing PyNaCl and developing secure execution environments for smart contracts with comprehensive transaction validation.

  • Blockchain
  • P2P
  • libp2p
  • Security

• Developer Intern Hyderabad

  at Saral ai

  MAY 2026 - JULY 2026

  • Architected automated model evaluation pipelines using LangSmith to benchmark, log, and monitor token latency, correctness, and performance across Gemini and Sarvam AI models.
  • Engineered context-aware LLM agent orchestration logic and shipped a production-ready TTS pipeline using asynchronous multi-threaded FFmpeg audio stitching.

  • LangSmith
  • LLM Evals
  • Gemini
  • TTS
  • FFmpeg

projects.

• Skillcheck

  github

  • Evidence-backed static analysis scanner for AI-agent skill packages, categorizing findings as MALICIOUS, DANGEROUS, SUSPICIOUS, UNVERIFIED, or NO_FINDINGS, each backed by quoted evidence.
  • Parses Markdown structure, recursively decodes obfuscated payloads (base64/hex/rot13), and runs detectors across prose, Python AST, shell commands, secrets, and Unicode tricks — with multilingual prompt-injection detection across 8 languages.

  • FastAPI
  • SQLite
  • Docker
  • Security
  • Static Analysis

• Slm From Scratch

  Confidential / Internal

  • Built a modular small language model architecture from the ground up, implementing Grouped Query Attention (GQA), SwiGLU activation functions, Rotary Positional Embeddings (RoPE), and RMSNorm layers.
  • Engineered custom variants including a compact 4M-parameter token generator and a Qwen2.5-micro replication to study compute-efficient model convergence constraints.

  • PyTorch
  • Python
  • Transformers
  • LLM

• Transformers From Scratch

  Confidential / Internal

  • Developed the complete vanilla Transformer architecture ("Attention Is All You Need") from raw mathematical specifications, without relying on high-level transformer packages.
  • Built custom scaled dot-product attention units, token embedding steps, static positional encodings, and optimization loops to analyze encoder-decoder layer mechanics during training.

  • PyTorch
  • NumPy
  • Deep Learning

technical blogs.