Skip to content
View SiddardhaShayini's full-sized avatar

Block or report SiddardhaShayini

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
SiddardhaShayini/README.md

Hi, I'm Siddardha Shayini 👋

AI/ML Engineer · building safe, local-first AI agents · creator of Actentra, OlaiEdge, TAEL-Agent, HERM4AI and StrataSLM · applied AI that holds up under real-world constraints, on devices never meant to run AI.

📍 Hyderabad, India  |  🟢 Open to AI/ML and software engineering roles

Portfolio LinkedIn Hugging Face Google Scholar ORCID


🧠 About Me

  • 💼 Role: AI/ML Engineer specializing in Agentic AI, AI agent security, Generative AI, LLM fine-tuning, agent memory, Computer Vision, and edge/local LLM inference.
  • 🔬 Focus: Local-first, privacy-preserving AI: deterministic security boundaries for agent tool calls, fine-tuning small models (QLoRA), serving them as GGUF on CPU-only hardware, and wrapping them in reliable agent runtimes (tool calling, persistent memory, context management).
  • 🛠️ Currently building: Actentra, a local-first security layer for AI agents.
  • 🎓 Education: B.Tech in CSE (AI & ML), Vignan Institute of Technology and Science, Hyderabad (2022–2026).
  • 👀 Interests: Applied AI, Agentic AI, AI agent security, Local LLMs, NLP, Computer Vision, MLOps.

🚀 Featured Projects

Project What it is Stack
🛡️ Actentra (currently building · Open Source) Lightweight, local-first deterministic execution boundary for AI agents. Sits between any model or agent framework and real tools (files, databases, HTTP): canonicalizes every call, evaluates versioned YAML policies and risk tags, binds human approvals to an exact sha256 action hash, executes with least privilege, and writes tamper-evident, hash-chained audit receipts, with no LLM in the enforcement path. Adds persistent single-use approvals, session taint tracking against indirect prompt injection, cumulative session budgets, SSRF-defended scoped executors, an MCP proxy with schema pinning, Ed25519-signed checkpoints, a policy canary diff engine, an offline operator console, and LangChain / OpenAI function-calling adapters. Replays 32 documented agent incidents. Python, YAML policies, SQLite WAL, MCP, Ed25519, Typer CLI
⚡ OlaiEdge (formerly FastAgent · Open Source, MIT) Headless, drop-in local LLM runtime (SDK + background daemon) for desktop apps. Cuts steady-state prompt latency from ~3,500 ms to ~73 ms on a CPU-only, 8 GB machine with zero OOM failures in long tool loops. Adds a RAM-budget enforcer, process watchdog, multi-tab slot cache, and offline signed licensing; 26 tests. Paper Python, llama.cpp, LCP slot caching, 8-bit KV-cache quantization, GBNF, Electron/Tauri/Flutter SDKs
🛠️ TAEL-Agent · 🤗 Model Privacy-preserving local autonomous SWE agent. IBM Granite 3.1 2B Instruct fine-tuned with QLoRA on a 29,900-sample, three-phase curriculum, deployed as a 4-bit GGUF. Runtime has 9-category task routing, 10 sandboxed workspace tools, four-pillar persistent memory, and context-budget management. VS Code extension with refactoring, diff previews, and autonomous tool execution. 8/8 benchmark-domain passes in the reported evaluation suite. QLoRA, GGUF, Ollama, TypeScript, VS Code API
🧠 HERM4AI (Open Source, MIT) Human-like Evolving & Reasoning Memory for AI: a model-free cognitive memory architecture for LLM agents (valid-time store, 7-slot working memory, associative knowledge graph, sleep-phase consolidation, ACT-R arbitration, Q-MIA provenance). Shows flat vector RAG suffers temporal inversion (41.7% obsolete retrievals), dual collapse (9.0% top-1 at 10k passages), and prompt poisoning (33–40% hallucination); HERM4AI reaches 100% on 7 benchmark suites at <2 ms CPU latency; 36 tests. Paper Python, SQLite, Knowledge Graphs, ACT-R
🐾 WildGuardAI (IEEE Xplore, 2026) Real-time wildlife-conflict prediction from camera feeds with GPS-based alerts for forest officers and villagers. YOLOv12n, BoT-SORT, Kalman Filter, Flask, SQLite

More projects:

  • 🧱 StrataSLM: featherweight, deterministic local AI runtime and bi-temporal cognitive memory engine for 8 GB, CPU-only laptops (prefix-canonical caching, 6-phase injection-safe agent state machine, CoALA 4-tier memory with hybrid RRF retrieval, 3-stage resource governor).
  • 🧓 JivanMitra: offline-first, voice-enabled, multilingual AI companion for elders in India (LangChain, RAG, agentic AI).
  • 🪶 Hearthwren (in development): local-first, CPU-only AI coding agent.
  • 🌍 Terrascope AI (live): zero-paid-API platform scoring any city or coordinate across 16 socio-environmental dimensions and 5 personas, using rule-based scoring over Open-Meteo, OpenStreetMap Overpass, and Wikipedia data. Deployed on Cloudflare Pages.
  • 🔒 Personal AI Agent: local-first, privacy-focused assistant (Ollama, LanceDB).
  • 📚 Offline Deep Research Agent: reasons over documents entirely offline (Ollama, Mistral, LangChain).
  • 🧑‍💼 TalentRank AI (live demo): AI-assisted recruiting platform with JWT auth, AI candidate ranking, and analytics (React 18, Vite, Express, SQLite).

Full list in the portfolio archive.


📄 Research

  • Beyond Flat Vector Retrieval: Mitigating the Dual Collapse and In-Context Prompt Poisoning of RAG via Structured Valid-Time Cognitive Architectures (HERM4AI) — Zenodo, 2026 · Read the paper
  • Hardware-Adaptive Inference Orchestration: Zero-Overhead Local LLM Agents on Resource-Constrained Edge Devices (OlaiEdge) — Zenodo, 2026 · Read the paper
  • Wild Guard AI: Real-Time Wildlife Conflict Prediction — IEEE Xplore, ICICI 2026 · View on IEEE Xplore

🤝 Open Source

  • Actentra, OlaiEdge, TAEL-Agent, HERM4AI and StrataSLM: Creator and maintainer.
  • sklearn-genetic-opt: Contributor. Documented parallel processing with n_jobs in GASearchCV (PR #178 merged, resolving Issue #152).

🛠️ Tech Stack & Skills

Area Tools
Languages Python, TypeScript, JavaScript, Java, C, C++, SQL
AI/ML & Vision LLMs, Agentic AI, LLM Fine-Tuning (QLoRA/LoRA), RAG, NLP, Computer Vision, scikit-learn, TensorFlow, YOLOv12n, BoT-SORT, Kalman Filter, Vector Embeddings, Knowledge Graphs, Hybrid Retrieval (BM25 + Vector, RRF), Bi-Temporal Memory
LLM & Agent Systems Ollama, llama.cpp, GGUF, KV-Cache Quantization, Prompt Caching, Grammar-Constrained Decoding (GBNF), Tool Calling, Function Calling, Persistent Agent Memory, ReAct, Deterministic Agent Runtimes, OpenAI-Compatible APIs
AI Agent Security Policy-as-Code (YAML), Prompt-Injection Defense, Taint Tracking, Least-Privilege Execution, Human-in-the-Loop Approvals, Tamper-Evident Audit Logs, Model Context Protocol (MCP), SSRF Defense, Ed25519 Signing, Fail-Closed Design
Frameworks & Tools FastAPI, Flask, LangChain, LangGraph, React, Node.js, Streamlit, LanceDB, ChromaDB, SQLite (WAL / FTS5 / sqlite-vec), ONNX Runtime, asyncio, Typer, pytest / Hypothesis, Ruff / Mypy, GitHub Actions, uv, Hugging Face, Git/GitHub, VS Code Extension Development, REST APIs

🏅 Certifications & Training

  • Oracle Certified Foundations Associate: Agentic AI
  • Deep Research with LangGraph (LangChain Academy)
  • Hugging Face Agents Course
  • Databricks Fundamentals (Databricks Academy)
  • TCS iON NQT (IT): Qualified

📫 Get in Touch

Open to AI/ML engineering and software roles: LinkedIn · Email · Portfolio


⚡ Fun Fact

I enjoy turning complex AI ideas into simple, usable systems, especially for people in low-resource settings and for devices that were never meant to run AI.

Pinned Loading

  1. TAEL_Agent TAEL_Agent Public

    An autonomous, local software engineering AI agent engineered for 8GB RAM consumer laptops without dedicated GPUs.

    Python 1

  2. Optimizing-Local-AI-Agents Optimizing-Local-AI-Agents Public

    Python 1

  3. Sklearn-genetic-opt Sklearn-genetic-opt Public

    Forked from rodrigo-arenas/Sklearn-genetic-opt

    ML hyperparameters tuning and features selection, using evolutionary algorithms.

    Python 1

  4. WildGuardAI-Real-Time-Wildlife-Conflict-Prediction WildGuardAI-Real-Time-Wildlife-Conflict-Prediction Public

    Jupyter Notebook 1 1

  5. HERM4AI HERM4AI Public

    Human-inspired, Explainable, Reconfigurable Memory for Artificial Intelligence.

    Python 1

  6. OlaiEdge OlaiEdge Public

    OlaiEdge (pronounced Oh-lai Edge, standing for Optimizing Local AI at the Edge) is a lightweight, headless inference and agent engine built specifically for desktop software companies (code editors…

    Python 1