My favourite bug was in someone else's data. While building a brain-signal benchmark on a public EEG dataset, I found eight "different" participants whose recordings were byte-for-byte identical, so every model trained on it was quietly being tested on data it had already seen. I published an audit tool so others can check their copies.
That is roughly how I work: build the system, then try hard to break it before anyone else does. I'm a B.Tech CSE (Data Science) student at MIT Bengaluru (2023–2027) working across LLM agents, retrieval, graph machine learning, generative models, forecasting and MLOps, and I like problems where mathematics, messy data and engineering meet.
kavya = {
"building": ["agentic LLM systems with human sign-off", "production ML pipelines"],
"researching": ["graph & temporal ML", "topology of brain networks", "uncertainty you can trust"],
"published": "MOGA-FCM @ CIMA 2026, NIT Puducherry",
"rule": "no claim without a baseline, a held-out test and a negative control",
}|
PalletCast · Forecasting & MLOps |
DelegNet · Security ML |
|
PIRGE-CodeGuard · Agent safety |
ClearDue · Agentic AI |
|
Cross-task EEG workload decoding · Brain AI |
INTERVENE · Graph ML for biology |
🤖 LLMs, RAG & agentic AI
| Project | What it does |
|---|---|
| pirge-codeguard | Pre-execution safety gateway for coding agents |
| cleardue | LangGraph agents for B2B short-payment investigation, with a verifier model |
| chaos-engineering | chaosctl: nine LangGraph agents run bounded, reversible chaos experiments with verified rollback |
| enterprise-knowledge-copilot | Governed hybrid-RAG copilot on Azure AI Search with citations and groundedness checks |
| revenueops-agent-control-tower | Multi-agent CRM revenue-risk detection with human-review routing |
| supportops-copilot | LLM issue triage that can only suggest; a person approves every action |
| chnage-shield | ChangeShield: self-repairing Playwright automations with human approval |
| context-aware-rag-evaluation | Dependency-free RAG harness: Hit@K, MRR, groundedness |
| ai-gtm-lead-enrichment-automation | Lead scoring and LLM outreach prompts with CRM-ready payloads |
📈 Forecasting, MLOps & reliability
| Project | What it does |
|---|---|
| palletcast-forecasting-mlops | End-to-end MLOps for 14-day pallet forecasting |
| carbon-twin-R | Building-energy digital twin: sequential fault detection, conformal intervals, fault localisation |
| model-calibration-diagnostics | ECE, Brier, reliability diagrams and temperature scaling under shift |
| cross-domain-visual-robustness | How classical models break from MNIST to SVHN |
🕸️ Graph ML, network science & applied mathematics
| Project | What it does |
|---|---|
| transferguard-curvature-audited-localization | Tracing attacks back to their source in evolving networks |
| graph-spectral-early-warning | When does a network's spectrum warn of a tipping point? 15,200-setting identifiability study |
| nwcf-eeg-topology | Clique filtration + persistent homology for brain networks, with stability proofs and an honest negative result |
| degree-collision-reconstruction-bound | Network archaeology: how much growth history a snapshot can reveal |
| thermodynamic-pa-inference-single-snapshot | Inferring preferential-attachment parameters from one snapshot |
| intervene-luad-graph-attention | Graph-sparse attention for lung-cancer relapse |
| qcertroute-quantum-expert-routing | Shot-adaptive quantum expert routing with certified top-ten decisions |
🔐 Security ML
| Project | What it does |
|---|---|
| DelegNet | Pre-signature detection of malicious EIP-7702 delegations |
| cost-aware-active-telemetry | Finding the first compromised host when telemetry has a price (LANL) |
🧠 Brain, health & physiological AI
| Project | What it does |
|---|---|
| eeg-load-signature | Cross-task EEG mental-workload decoding |
| cogbci-integrity-audit | Found 8 COG-BCI participant IDs with identical recordings |
| hcp-meg-brain-state-pilot | MEG source-space pipeline and three pre-registered no-go probes |
| moga-fcm-cognitive-load-detection | Unsupervised driver cognitive-load detection (CIMA 2026) |
| gluco-adapt-healthcare-ai | Risk-aware adaptive sensing for wearable glucose estimation |
| hart-sampling-rate-adapters | Where to place adapters when a wearable sensor samples slower |
🎨 Generative models & Earth observation
| Project | What it does |
|---|---|
| curvature-constrained-latent-diffusion | Latent diffusion that respects curvature and topology on real Kvasir-SEG masks |
| Methane-twin-R | Physics-informed detection and quantification of methane plumes from NASA EMIT |
MOGA-FCM: Unsupervised Cognitive Load Detection via Multi-Objective Genetic Algorithms and Bayesian Fuzzy C-Means in ADAS
K. Kalavala, A. Pandey, S. Priya · 2nd Congress on Intelligent Machines and Algorithms (CIMA 2026), NIT Puducherry
Open to AI/ML engineering roles and research collaborations.
Every repository here has a README that says what worked, what didn't, and how to reproduce it.