B.Tech CSE (AI & ML) at SGT University, GPA 8.8/10. Built Katha-LM, a Hindi transformer LM trained from scratch, and Artha, a deployed Hinglish sentiment-analysis API; lead developer of ReVive, an AI physiotherapy platform incubated under IHFC, IIT Delhi.
Python · PyTorch · Custom BPE Tokenizer · FastAPI · Docker · Next.js
A decoder-only transformer built from scratch in PyTorch — bigram → MLP → transformer (val perplexity 26.8 → 16.3 → 13.1) — then scaled to a 5.19M-parameter config at val perplexity 2.51 on 100k Hindi stories (Kaggle T4). Engineered a custom Hindi BPE tokenizer (vocab 756) at 1.63 fertility vs GPT-2's 7.02 on Devanagari (~4× fewer tokens/word), shipped via FastAPI on HF Spaces (Docker) + Next.js 14 on Vercel.
Python · PyTorch · xlm-roberta · HuggingFace · FastAPI · Docker · Next.js
Scraped and processed 300k+ real Hinglish comments, applied domain-adaptive pretraining (DAPT) on xlm-roberta-base, then a 3-stage fine-tune (DAPT → v1 F1 0.7856 → neutral upsampling + lr 5e-6 → v2 Macro F1 0.9304). Deployed as a production REST API: FastAPI on HF Spaces (Docker) + Next.js on Vercel, with bearer-token auth, model versioning, and private HF Hub model hosting.
n8n · Gemini (Flash-Lite / Flash) · LlamaParse · Playwright · Telegram · Gmail · Google Sheets
An n8n agent that turns inbound RFQs (PDF / Excel / email) into branded PDF quotes end-to-end: LLM line-item extraction → in-context catalog matching & pricing over 75 SKUs → Playwright PDF → Telegram human-approval gate → in-thread Gmail reply + Google Sheets logging. Routed extraction and scoring to Gemini Flash-Lite and drafting to Flash, with strict structured-output schemas and confidence-based exception flagging so low-confidence matches surface for review instead of failing silently.
Python · LangGraph · LangChain · Chroma · RAGAS · Langfuse · FastAPI
A self-correcting RAG agent (LangGraph) over 491K peer-reviewed clinical references that grades its own retrieved evidence, rewrites weak queries before answering, and refuses what it can't ground — every claim cited. Benchmarked on the academic MIRAGE medical-QA benchmark at a 62.2% exam-task mean, with a controlled embedding ablation (general vs domain-specific MedCPT) that surfaced a retriever/reranker stage-misalignment: the domain model retrieved more but scored −4.4 pts lower downstream. Includes a resumable RAGAS / DeepEval eval harness with Langfuse tracing.
React Native · Node.js/Express · MongoDB · Redis · Socket.IO · WebRTC · IHFC, IIT Delhi
Lead developer of a cross-platform patient + therapist app: 7-step clinical onboarding, real-time booking against live availability, and in-session WebRTC video consultations over a Socket.IO signalling namespace. Engineered the booking engine with 30-minute slot generation and a 3-layer double-booking guard (Redis lock + DB uniqueness constraint + partial-index backstop).
Lead Developer & Co-Founder
Lead developer and co-founder of an AI physiotherapy platform incubated under IHFC, IIT Delhi — own the architecture and delivery of the cross-platform patient + therapist app end-to-end.
Event Coordinator
Planned and executed 5+ public-speaking & debate events end-to-end; managed logistics and multi-channel outreach.
B.Tech — Computer Science & Engineering (AI–ML Specialization) · GPA 8.8 / 10