Open to AI/ML engineering roles · Hyderabad, IN

Siddardha Shayini
builds systems that hold up.

I'm an AI/ML engineer and the creator of OlaiEdge (formerly FastAgent), an open-source local LLM runtime that cut prompt latency from ~3,500 ms to ~73 ms on 8 GB CPU-only hardware with zero OOM crashes; of HERM4AI, a model-free cognitive memory architecture for LLM agents; and of TAEL-Agent, a privacy-preserving autonomous software-engineering agent built on a QLoRA fine-tuned 2B model that runs fully on-device. I also work in computer vision and generative AI, including a real-time wildlife-conflict detection system published in IEEE Xplore. I care more about whether a system holds up under real conditions than whether the demo looks good.

Download Resume View Case Files Read the OlaiEdge Paper
GitHub GitHub LinkedIn LinkedIn IEEE Xplore IEEE Xplore Google Scholar Google Scholar ORCID ORCID

30 projects logged
3 publications (IEEE Xplore + 2 Zenodo)
6 internships
1 merged open-source PR
Siddardha Shayini
About

AI/ML engineer, founder, open-source maintainer

I graduated with a B.Tech in CSE (AI & ML) from Vignan Institute of Technology and Science, Hyderabad, and I'm looking for my first full-time role in AI/ML and software engineering. Right now I'm building TAEL-Agent, a local autonomous SWE agent: IBM Granite 3.1 2B fine-tuned with QLoRA, deployed as a 4-bit GGUF, wrapped in a sandboxed tool runtime with persistent memory and a VS Code extension, all on CPU-only, 8 GB machines. I also created and maintain OlaiEdge (formerly FastAgent), a drop-in local LLM runtime for desktop apps, and published HERM4AI, an open-source cognitive memory architecture for LLM agents that avoids the temporal inversion and prompt-poisoning failures of flat vector RAG. Both have papers on Zenodo.

Before that, WildGuardAI predicted human–wildlife conflict in real time from camera-trap-style footage and was published in IEEE Xplore. I've also shipped full-stack products such as Terrascope AI (a zero-paid-API city intelligence platform) and TalentRank AI. Most of what I build works within real constraints: JivanMitra, my personal AI agent, and my research-agent projects run local-first and offline, because I'm more interested in systems that hold up without a safety net than ones that only work with an API key and a stable connection.

Field Record

Experience

NOW

Founder & AI Engineer — TAEL-Agent

Sep 2026 – Present · Open Source

Built a privacy-preserving local autonomous SWE agent for CPU-only, 8 GB RAM systems, fine-tuning IBM Granite 3.1 2B Instruct with QLoRA on a 29,900-sample, three-phase curriculum and deploying a 4-bit GGUF model locally. Engineered an autonomous runtime with 9-category task routing, 10 sandboxed workspace tools, four-pillar persistent memory, and context-budget management. Developed a VS Code extension with code generation, refactoring, diff previews, workspace-aware assistance, and autonomous tool execution; achieved 8/8 benchmark-domain passes in the reported evaluation suite.

GitHub → Hugging Face →
NOW

Creator & Open Source Maintainer — OlaiEdge (formerly FastAgent)

Jun 2026 – Present · Open Source · MIT

Built a headless, drop-in local LLM runtime (SDK + background daemon) on llama.cpp using Longest-Common-Prefix slot caching, 8-bit KV-cache quantization, GBNF grammar-constrained JSON, and a RAM-budget enforcer for desktop apps (Electron/Tauri/Flutter SDKs). Reduced steady-state prompt-evaluation latency from ~3,500 ms to ~73 ms on a CPU-only, 8 GB machine with zero OOM failures in long tool loops; added a process watchdog, multi-tab slot cache, and offline signed licensing; 26 tests.

GitHub → Paper (Zenodo) →
RES

Creator & Researcher — HERM4AI

Sep 2026 – Nov 2026 · Open Source · MIT

Designed an open-source, model-free cognitive memory architecture for LLM agents: valid-time store, 7-slot working memory, associative knowledge graph, sleep-phase consolidation, ACT-R arbitration, and Q-MIA provenance; the core engine has zero neural dependencies. Showed that 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 reached 100% on 7 benchmark suites at <2 ms CPU latency, with 36 passing tests.

GitHub → Paper (Zenodo) →
LOG-06

Machine Learning & AI Intern — Skill Nexis

Aug 2026 – Sep 2026 · Remote

Worked on machine learning models and related hands-on AI/ML tasks as part of a remote internship.

OSS

Open Source Contributor — sklearn-genetic-opt

Sep 2025 – Mar 2026 · Open Source

Documented parallel processing with n_jobs in GASearchCV. Pull request #178 merged, resolving Issue #152.

View pull request →
EDU

B.Tech, CSE (AI & ML)

Vignan Institute of Technology and Science, Hyderabad · 2022–2026 · CGPA 7.72/10.00
LOG-01

AIML Intern — Elevate Labs

May 2025 – Jun 2025 · Remote

Contributed to hands-on AI/ML projects and tasks, building practical experience across the model development workflow.

LOG-02

Web Development Intern — CollegeTips.in

Apr 2025 – May 2025 · Remote

Built interactive web platform features and visual content. Researched AI-driven ideas for recruitment tools and chatbots, and wrote reports and content for the platform.

LOG-03

AI Engineer Intern — Infosys Springboard

Feb 2025 – Apr 2025 · Remote

Built a smart electricity demand & price forecasting system using Facebook Prophet time-series models, reaching ~95% accuracy on real market datasets. Developed the Flask backend and a Tailwind-styled UI for dynamic date-range inputs with real-time forecast visualization.

LOG-04

AI Engineer Intern — YBI Foundation

Dec 2024 – Jun 2025 · Remote

Developed an Air Quality Prediction System for Indian cities using Random Forest Regression on PM2.5, PM10, and NOx pollutant data, reaching ~96% AQI category accuracy. Built interactive CLI visualizations and data pipelines across multiple city datasets.

LOG-05

Python Developer — YBI Foundation

Oct 2024 · Remote

Built a movie recommendation engine suggesting personalized picks based on user preferences and historical data — the earliest version of what's now the Movie Recommendation System project in the archive below.

Instrumentation

Skills

Languages
PythonTypeScriptJavaScript JavaCC++HTML/CSS
AI / ML
LLMsAgentic AI LLM Fine-Tuning (QLoRA/LoRA)RAGNLP Computer Visionscikit-learnTensorFlow YOLOv12nBoT-SORTKalman Filter Vector EmbeddingsKnowledge GraphsTime-Series Forecasting
LLM & Agent Systems
Ollamallama.cppGGUF KV-Cache QuantizationPrompt Caching Grammar-Constrained DecodingTool Calling Persistent Agent MemoryReAct Agents
Frameworks & Tools
FastAPIFlaskLangChainLangGraph ReactNode.jsStreamlit LanceDBChromaDBSQLite Hugging FaceGit / GitHub VS Code Extension DevelopmentREST APIs
Other
Data AnalysisDatabricks Fundamentals Deep LearningOpen Source MaintenanceFlutter (basics)
Certifications & Training
Oracle Certified Foundations Associate — Agentic AI Deep Research with LangGraph — LangChain Academy Hugging Face Agents Course Databricks Fundamentals — Databricks Academy NESTGEN E-Learning Festival 2026 TCS iON NQT (IT) — Qualified NPTEL — Programming with Generative AI Coursera — Cloud ML Engineering & MLOps
Case Files

Featured Work

LOG-00 · Jun 2026–Present Open Source · Actively Maintained

OlaiEdge formerly FastAgent

A headless, drop-in local LLM runtime (SDK + background daemon) for desktop apps. Naive local agent loops re-evaluate the system prompt every turn and let context grow until memory runs out. OlaiEdge wraps llama.cpp with Longest-Common-Prefix slot caching, 8-bit KV-cache quantization, GBNF grammar-constrained JSON, and a RAM-budget enforcer, plus a process watchdog, multi-tab slot cache, and offline signed licensing. Ships with Electron, Tauri, and Flutter SDKs and 26 tests.

Pythonllama.cppKV-Cache Quantization Prefix CachingGBNFElectron / Tauri / FlutterMIT License
Explore on GitHub → Read the paper →
Benchmark · CPU-only, 8 GB RAM
Prompt eval latency~3,500 ms~73 ms
KV-cache precisionf16q8_0
Long tool loopsOOM0 OOM failures
Test suite26 tests
LOG-00B · Sep–Nov 2026 Open Source · Research

HERM4AI

Human-like Evolving & Reasoning Memory for AI: a model-free cognitive memory architecture for LLM agents with a valid-time store, 7-slot working memory, associative knowledge graph, sleep-phase consolidation, ACT-R arbitration, and Q-MIA provenance. Shows flat vector RAG fails on temporal inversion (41.7% obsolete retrievals), dual collapse (9.0% top-1 at 10k passages), and prompt poisoning (33–40% hallucination); HERM4AI hits 100% on 7 benchmark suites at <2 ms CPU latency, with zero neural dependencies in the core engine and 36 passing tests.

PythonSQLiteKnowledge GraphsACT-RMIT License
GitHub → Paper (Zenodo) →
LOG-00C · Sep 2026–Present Open Source · Local-First

TAEL-Agent

A privacy-preserving autonomous SWE agent for CPU-only, 8 GB machines. IBM Granite 3.1 2B Instruct fine-tuned with QLoRA on a 29,900-sample, three-phase curriculum and deployed as a 4-bit GGUF. The runtime adds 9-category task routing, 10 sandboxed workspace tools, four-pillar persistent memory, and context-budget management, surfaced through a VS Code extension with diff previews and autonomous tool execution. Passed 8/8 benchmark domains in the reported evaluation suite.

QLoRAIBM Granite 3.1GGUF VS Code ExtensionAgent Memory
GitHub → Hugging Face →
LOG-01 · Aug 2025–Jun 2026 IEEE Xplore · Published

WildGuardAI

Deep learning system using YOLOv12n, BoT-SORT, and a Kalman Filter to detect animals crossing forest boundaries from live camera feeds. Flask backend with SQLite and GPS-based alerting notifies forest officers and villagers before dangerous encounters happen. Published in IEEE Xplore as "Wild Guard AI: Real-Time Wildlife Conflict Prediction" (2026), and presented at ICICI 2026.

YOLOv12n BoT-SORT Kalman Filter Flask SQLite
Explore on GitHub →
LOG-02 · Jul 2025–Present Active Development

JivanMitra

Offline-first, voice-enabled AI companion for elderly users in India. Uses LangChain RAG, agentic AI, LLMs, and NLP to hold context-aware, multilingual conversations in local languages, without depending on a stable internet connection — built for accessibility, not tech fluency.

LangChain RAG Agentic AI NLP Multilingual
Explore on GitHub →
Full Archive

25 More Projects

GenAI & NLP

Hearthwren in development

Local-first, CPU-only AI coding agent (CLI: hwren), currently in active development.

PythonLocal-FirstCoding Agent
GitHub →
Web & Apps

Terrascope AI

Zero-paid-API platform scoring any city or coordinate across 16 socio-environmental dimensions and 5 personas (family, digital nomad, tourist, retiree, career). Deterministic rule-based scoring over Open-Meteo, OpenStreetMap Overpass, and Wikipedia REST data with a WAL-mode SQLite cache; deployed on Cloudflare Pages.

Rule-Based AIOpen-MeteoOSM OverpassSQLite
Live →
GenAI & NLP

Personal AI Agent — Local-First & Private

A local-first, privacy-focused AI assistant built with Ollama and LanceDB, keeping all data and inference on the user's machine.

OllamaLanceDBLocal-First
Explore →
GenAI & NLP

Offline Deep Research Agent

A research agent that reads and reasons over documents entirely offline. Rebuilt on Ollama and Mistral after the original API-dependent version proved too fragile for reliable local use.

OllamaMistralLangChain
Explore →
Web & Apps

TalentRank AI

AI-assisted recruiting platform: JWT auth, job and candidate management, resume upload, AI scoring and ranking of applicants, hiring analytics dashboard, and recruiter activity feed. React 18 + Vite frontend, Express.js + SQLite backend, deployed on Vercel and Render.

ReactExpressSQLiteJWT
Live → Code →
GenAI & NLP

Personal Research Assistant — Offline Agentic AI

Offline research assistant using RAG, local embeddings, ChromaDB, and Ollama Phi-3 to analyze PDFs with multi-step reasoning.

RAGChromaDBOllama
Explore →
GenAI & NLP

JivanMitra: Your AI Elder Companion

An earlier, voice-first iteration of JivanMitra designed to support elderly users in India.

Voice AILangChain
Explore →
Web & Apps

Voice-Controlled Virtual Wheelchair

Hands-free navigation through voice commands in an immersive 3D browser environment, built for accessibility.

Web Speech API3D
Explore →
Data & ML

Visual EEG Pattern Synthesis & Recognition

Analyzes and generates EEG signal patterns using deep learning techniques.

Deep LearningSignal Processing
Explore →
Computer Vision

2D to 3D Converter

Converts photos or text prompts into downloadable 3D models using open-source AI/ML libraries.

Generative 3DStreamlit
Explore →
Web & Apps

Pokemon Explorer

Search, filter, and browse Pokémon by type and detail using the PokéAPI.

StreamlitREST API
Explore →
Data & ML

Electricity Demand & Price Forecasting

Predicts future electricity demand and price from historical data through a modern interface.

ForecastingML
Explore →
Data & ML

Cyber Crime Forecasting & Analysis in India

Streamlit dashboard forecasting cybercrime trends across Indian states using ARIMA and Polynomial Regression, reaching ~93% accuracy, to support awareness and policy planning.

ARIMAPolynomial RegressionStreamlit
Explore →
Data & ML

Smart Weather Advisor

Personalized recommendations based on real-time weather conditions to help users make informed decisions.

Weather APIStreamlit
Explore →
GenAI & NLP

Digital Literacy Companion

A chatbot designed to help parents and elderly users navigate the digital world with confidence.

ChatbotStreamlit
Explore →
Data & ML

Classification Algorithms on Heart Disease Indicators

Evaluates and compares classification algorithms on the Indicators of Heart Disease dataset.

Classificationscikit-learn
Explore →
Data & ML

Data Analysis & ML Algorithms in Python

A reference implementation of core data analysis and machine learning algorithms in Python.

PythonML
Explore →
Data & ML

Movie Recommendation System

Suggests movies based on user preferences using the Nearest Neighbors algorithm.

Recommender Systems
Explore →
Web & Apps

FashionMate

Showcases fashion trends and lets users create personal style boards.

Web App
Explore →
Data & ML

Medical Data AI — Vault

A scalable graph database system for secure healthcare data management with AI-powered insights.

Graph DBHealthcare
Explore →
Web & Apps

Occasion-specific Stylist AI

Personalized fashion recommendations based on event details.

Recommendation
Explore →
Web & Apps

Expense Tracker

An easy-to-use income and expense tracking app.

Web App
Explore →
Web & Apps

Interior Design Showcase

Browse stylish interior design ideas to inspire and enhance living spaces.

Web App
Explore →
Web & Apps

Flashcard App

A flashcard app for efficient learning, with customizable cards and export options.

Web App
Explore →
Web & Apps

Flutter Experiments

Flutter concepts including animations, API integration, and responsive design.

FlutterDart
Explore →
Outside of engineering, I write — including Life's Odyssey, a personal reflection on growth and legacy.
Verification

Research & Profiles

Preprint · Zenodo · Nov 2026 · MIT

Beyond Flat Vector Retrieval: Mitigating the Dual Collapse and In-Context Prompt Poisoning of RAG via Structured Valid-Time Cognitive Architectures

A model-free cognitive memory architecture for LLM agents (valid-time store, working memory, associative knowledge graph, sleep-phase consolidation, ACT-R arbitration, Q-MIA provenance). Shows flat vector RAG suffers temporal inversion, dual collapse, and prompt poisoning; HERM4AI reaches 100% on 7 benchmark suites at <2 ms CPU latency.

Read on Zenodo → DOI: 10.5281/zenodo.23978208 Source code →
Preprint · Zenodo · Sep 2026 · CC BY 4.0

Hardware-Adaptive Inference Orchestration: Zero-Overhead Local LLM Agents on Resource-Constrained Edge Devices

Presents OlaiEdge (formerly FastAgent), a headless inference orchestrator combining LCP slot caching, 8-bit KV-cache quantization, grammar-constrained decoding, and an Adaptive Compute Controller. On a CPU-only 8 GB baseline, steady-state prompt-evaluation latency fell from ~3,500 ms to ~73 ms with no OOM failures during long-running tool loops.

Read on Zenodo → DOI: 10.5281/zenodo.22918273 Source code →
IEEE Xplore · ICICI 2026

Wild Guard AI: Real-Time Wildlife Conflict Prediction

Real-time computer vision system using YOLOv12n, BoT-SORT, and a Kalman Filter to predict human–wildlife conflict from camera feeds, with GPS-based alerts for forest officers and villagers.

View on IEEE Xplore → Source code →
GitHub Hugging Face Google Scholar ORCID IEEE Xplore Semantic Scholar
Get In Touch

Contact

Gmail LinkedIn Resume ↗