📍 Mumbai, India | 🎓 GRE: 333/340 (V:165, Q:168) | ✉️ sahil.bhardwaj@somaiya.edu
I split my time between NLP research (cross-lingual transfer, LLM evaluation on non-Latin scripts) and building full production systems RAG platforms, ITSM tools, agent harnesses, and DeFi UIs. I like projects that go end-to-end: architecture, infra, and the paper/README to explain why it works.
I studied how model scale and prompting strategy affect LLM performance across diverse script families.
Current paper: Scale Helps, Few-Shot Hurts: Evaluating LLMs Across Five Non-Latin Scripts
- Scaling LLaMA 3.1 8B → 3.3 70B closes the English-bias for Hindi, Thai and Chinese (+8% each), but Arabic uniquely resists (+2%)
- Few-shot prompting improves English (+3%) but hurts every non-Latin script except Chinese — Arabic drops −6%
- Tokenizer inflation alone doesn't explain the Arabic gap — points to morphological complexity as the real driver
| Project | What it is | Stack |
|---|---|---|
| PromptLens | Observability stack for LLM apps and AI agents — captures prompts, responses, tool calls, memory ops, and RAG retrievals in a real-time, DevTools-style dashboard with session tracing and cost/latency analytics | FastAPI, TanStack Start (React), Python SDK, WebSocket, PostgreSQL |
| Granthiq | Self-hostable AI document intelligence platform (NotebookLM-style) - hybrid RAG, HyDE query expansion, Cohere reranking, hallucination guardrails, a research agent, and MCP tools for Cursor/Claude Desktop | Next.js 16, FastAPI, Qdrant, Supabase, LlamaIndex |
| nirmaan | Model-agnostic terminal coding agent, fully instrumented for research - event traces, cache-aware compaction, ablation toggles, and benchmark suites | Python, Groq/OpenAI/vLLM/Ollama, OpenTelemetry |
| sarvamdesk | Production-grade IT Service Management (ITSM) platform with microservices for tickets, assets, SLAs, and a knowledge base | Java 17, Spring Boot 3, React, Kafka, PostgreSQL |
| Frost | Solana staking platform - non-custodial staking UI with live TVL/APY stats and DeFi ecosystem integrations (Orca, Jupiter, Solend, Marinade) | React, TypeScript, Vite, Tailwind |
| Fusion | React + TypeScript frontend project (Vite, shadcn/ui) | React, TypeScript, Vite |
| llm-multilingual-evaluation | Research repo for the paper above - LLaMA 8B vs 70B across 5 non-Latin scripts on XQuAD | Python, Jupyter, Groq API |
- ML / NLP: HuggingFace, PyTorch, LLaMA, Groq API, RAG, XQuAD, LlamaIndex
- Backend: Python, Java (Spring Boot), FastAPI, Node.js , Express.js
- Frontend: React, Next.js, TypeScript, Tailwind CSS
- Infra / DevOps: Docker, Kubernetes, Tekton Pipelines, Terraform, CI/CD
- Data / Vector: PostgreSQL, Qdrant, Redis, Kafka
- Cloud: Google Cloud Platform (Certified), AWS, Azure
- Google Cloud Professional | Google | 2024–2026 |
- GRE: 333/340 (V:165, Q:168) | ETS | 2024 |
Tekton Pipelines - github.com/tektoncd/pipeline Contributed to the Kubernetes-native CI/CD pipeline project used across cloud-native production environments.
- Multilingual NLP and cross-lingual transfer
- LLM evaluation across low-resource languages
- Indic language NLP (Hindi, Bengali, Tamil)
- Efficient multilingual models and encoder injection