Explore my portfolio ↗ / Connect on LinkedIn ↗ / Explore the code ↓
I'm Ninad, an AI engineer building LLM applications, computer vision experiments, and systems for evaluating model reliability. I like the whole loop: formulate a question, build the system, inspect its failures, and make the results usable.
Currently building conversational AI, LangGraph workflows, and RAG services at ElevateSoul.ai as an AI Engineering Intern. My side quests tend to involve space.
Can a model read a document accurately and still miss its manipulation? A reproducible VLM/OCR benchmark with 600 synthetic documents, counterfactual pairs, hard negatives, and calibration analysis.
|
A paper-inspired study of ID presentation attacks using semantic, texture, and edge representations. Three model configurations, group-aware splits, and failure analysis on a controlled 300-sample protocol.
|
A physics-based Mars entry, descent, and landing simulation with a reinforcement-learning policy interface and a mission-control dashboard driven by Python environment telemetry.
|
Telecom troubleshooting through retrieval-augmented diagnostics: FAISS retrieval, cross-encoder reranking, structured LLM outputs, and explicit confidence and failure handling.
|
- Reading isn't verification. NirikshaID's Qwen2.5-VL run reached 98–99% field normalized exact match while recording 0% tamper F1 on two small, 20-sample evaluation splits. Results and limitations →
- Extra representations need controlled evaluation. IDSentinel's texture branch reached 0.679 macro-F1 versus 0.590 for its semantic baseline on a 44-image test split. A small, single-seed result with architecture differences—not a statistical-superiority claim. Protocol and analysis →
| Area | Tools I work with |
|---|---|
| LLM systems | Python · LangGraph · RAG · Transformers · FAISS |
| Vision & evaluation | PyTorch · OpenCV · scikit-learn · OCR · calibration · failure analysis |
| APIs & data | FastAPI · PostgreSQL · Supabase |
| Interfaces & automation | TypeScript · React / Next.js · Playwright · FFmpeg |
Experience & more things I've built
| Role | Focus |
|---|---|
| AI Engineering Intern · ElevateSoul.ai | Conversational flows, RAG services, and PostgreSQL/Supabase data workflows |
| AI Engineering Intern · Fanisin | Speech, LLM, memory, and media pipelines; Playwright and FFmpeg automation |
| AI Research Intern · Computer Society of India, Hyderabad Chapter | Transformer inference, controlled interventions, and token-level analysis |
| Data Science Intern · ZeTheta Algorithms | Representation learning and anomaly detection with autoencoders, PCA, and UMAP |
- LexiAudit — contract parsing, entity extraction, semantic retrieval, and rule-based risk analysis.
- GLaDOS — a cinematic conversational AI experience with voice synthesis and a browser demo mode.
- Gaia Latent Space Anomalies — unsupervised exploration of Gaia DR3 stellar populations.
- Astro ML Dashboard — photometric classification with uncertainty and out-of-distribution detection.
Have an interesting AI problem? Let's talk on LinkedIn, or take a look around my portfolio.
Visakhapatnam, India · Building useful systems. Asking better questions. Occasionally landing on Mars.