Machine Learning
Simulation-informed modelling, HVAC prediction, clustering, and explainability for the built environment.
Explore Live DemosApplied AI Systems
Multi-agent architecture, retrieval, local-model evaluation, and research tools that make complex model behaviour easier to inspect.
Browse Selected WorkResearch Communication
Interactive demos, dashboards, workshops, and evidence-aware stories that connect technical work with people.
View Talks & Events// 01
Talks & Events
Technical ideas made legible through workshops, talks, and community events.
Past talk · Speaker
Cracking the Formula for AI & ML
A practical map of AI beyond language models: supervised and unsupervised learning for structure, RAG for factual grounding, and agents for planning, memory, and action. A playful embedding-space example showed how relationships in meaning space can help explain one path to model hallucination.
Past workshop · Speaker
ML, LLMs, RAG & AI Agents
A company workshop covering data and ML fundamentals, transformer architecture, retrieval-augmented generation, and the profile–memory–planning–action model of AI agents. Teams finished by sketching real automation ideas, from sales workflows to office-fridge assistance.
Watch video ↗Featured video · Guest
How AI Could Help Buildings Manage Energy Demand
A conversation about the potential for AI to help buildings manage energy demand.
View event post ↗Past event · Speaker
AI Dev Day India 2026
Presented on local-LLM hallucination mitigation through agent comparison at a day of AI, quantum-computing, and real-world adoption discussions. Sessions also explored quantum systems, production AI-agent evaluation, and responsible AI practice.
View event post ↗Past event · Host
The AI Collective × Deck Meetup
Hosted with Deck and guest speaker Artem Avdacev on what makes a product AI-native. The discussion focused on building trust through small validated steps, then using context and memory systems to improve products over time.
View event post ↗Past event · Host
The Rise of Everyday Data Scientists & Prompt Engineers
A completed community event about how product teams, marketers, designers, founders, and operators are increasingly working with data and prompts—not only technical specialists. The LinkedIn recap captures the room, Sarra Kallel’s talk, and the event’s practical framework for clearer prompting.
// 02
Live Demos
Working experiments hosted on Hugging Face Spaces. Each card launches the real interface; a sleeping Space may take a moment to wake.
Concept visual
Building & Energy ML
Building Energy Predictor
Change weather and seasonal inputs to test an LSTM-based HVAC energy forecast. The interface exposes the scenario controls, predicted load, and recent trend so model behaviour can be explored rather than treated as a black box.
Hosted on Hugging Face · may wake on launch
Launch demo
Concept visual
Building Intelligence
Building Archetype Semantic Search
Describe a building in plain language—such as an old Montréal house or a BACnet office—and retrieve ranked matches from 100 archetypes. FAISS finds semantic candidates, SQLite returns the full records, and a metadata reranker explains why each result matched. This prototype is retrieval-only; it does not forecast or control buildings.
Hosted on Hugging Face · may wake on launch
Launch demo
Concept visual
Applied AI Systems
Multi-Agent System
Explore 199 stored agent responses across tasks and local models. Rotate a three-dimensional embedding space, compare clusters, inspect neighbouring responses, and see where agent strategies converge or diverge.
Hosted on Hugging Face · may wake on launch
Launch demo
Vector Memory
Semantic search across project knowledge using cosine and Euclidean similarity. This module is not yet a live demo.
// 03
Selected Work
Research systems, local AI infrastructure, data products, and visual research communication.
Multi-Agent LLM System
A local-model comparison and agent-routing framework evolved from hallucination-mitigation research. It stores responses in FAISS, studies semantic agreement with PCA and KMeans, and weighs task fit, consensus, runtime, and energy to favour the leanest capable model.
Concept visualAIOS: Local AI Operating System
A work-in-progress, privacy-first desktop automation layer. It captures screen context, extracts text and interface elements, asks a local or remote model to plan, and executes approved actions through input controls, APIs, or scripts with replayable logs and memory.
Concept visualAI Interconnect
The Apple-side node of AIOS: a SwiftUI app for macOS and iOS that unifies local MLX, Apple Intelligence, and Core ML backends, while connecting securely to the AIOS hub for devices, scenes, topology, and natural-language commands.
Concept visualStock Market Ensemble with Sentiment Analysis
A NIFTY 50 forecasting pipeline combining Prophet, logistic regression, random forest, technical indicators, India VIX, and VADER sentiment from financial headlines, followed by a risk-controlled strategy simulation.
Recruitment Database
A recruitment database design in MySQL, presented in Canva through its core entities, relationships, and database queries.
Dental Age Estimator
A machine-learning presentation exploring model evaluation for a dental-age estimation task.
View original presentation ↗
Information Security · Survey Research
Digital Footprints & Corporate Data Collection
How do the devices and services we use relate to the advertising we see? This exploratory study examines device ecosystems and people’s reported experience of targeted ads, opening a wider discussion about how major platforms collect and interpret everyday data.
How major platforms collect data
Data can come from information you provide, activity a service records, or other businesses. Companies can use those signals to estimate interests and personalize experiences. Collection and use vary by product, region, and privacy settings.
- Its policy describes collecting searches, video activity, device identifiers, and location information depending on services and settings. Cookies, pixels, and partner sites using Google services can also contribute activity data. Google describes using activity to personalize recommendations and ads.
- Google Privacy Policy ↗
- Apple
- Its policy lists account and device information, transactions, usage and diagnostic data, and location in specified circumstances. Apple describes service-specific collection and personalization choices, and states that its advertising platform does not link its data with third-party data for targeted advertising.
- Apple Privacy Policy ↗
- Meta
- Other businesses can share activity from their websites and services, such as purchases or games played. Meta describes using this information for ad personalization and expanding its use to other content, with availability and controls varying as changes roll out.
- Meta’s data-use explanation ↗
Inside the study
A Google Forms survey shared with friends and family supplied self-reported device ownership and advertising experiences. The workflow converted responses into binary features, defined ecosystem labels, handled missing values, and explored decision trees in RapidMiner, including SMOTE for class imbalance.
What the evidence supports
The survey explores patterns in reported experiences. It does not measure data captured inside a company or establish that an ecosystem causes more tracking. SMOTE changes class balance; it does not make a convenience sample representative.
Platform context draws on the linked company publications, reviewed October 28, 2026.
// Founder
Founder, Le Laboratoire Kalkin
Gauraang is a Building Engineering PhD researcher, McGill graduate, and founder of Le Laboratoire Kalkin.
The work connects intelligent systems, buildings, research practice, culture, music, and creative experimentation.
Evidence boundary: research claims require evidence; philosophy and creative inquiry offer questions, metaphors, and perspective rather than scientific proof.
// Book
Pick a time
Live availability straight from my calendar, so every slot you see is genuinely free. Meetings run on Google Meet, and times are shown in your own timezone.
Calendar not loading? Open the booking page directly, or email gauraangmalik1@gmail.com.