Final-year B.Tech student at IIT Roorkee and Applied Scientist Intern at Amazon, working on machine learning systems that are measured, reproducible, and production-shaped. My work spans LLM post-training and reasoning, agentic systems and LLMOps, and representation learning for both language and scientific data.

I care about the full loop — from a research idea to a rigorous evaluation harness to a containerized, deployable service — and I default to measuring claims rather than asserting them.

Interests

• LLM reasoning  ·  post-training (SFT, PEFT, distillation)
• Agentic memory  ·  multi-agent orchestration  ·  LLMOps / evaluation
• Representation & contrastive learning
• Information retrieval  ·  ML for scientific discovery

May 2026 – Aug 2026 · Bangalore  |  Representation learning · geospatial ML · uncertainty

• Trained a contrastive location embedding (CLIP-style InfoNCE with a frozen geo-encoder warm-start and time-conditioned fusion) that distills operational behavior into a coordinates-plus-time "fingerprint," so delivery-time predictions generalize to cold-start zones with no history.
• Engineered a reusable feature pipeline over 49M+ orders — traffic-friction speed, order density, and dwell time from delivery GPS — aggregated per zone × hour × day-of-week at multiple spatial scales.
• Built a rigorous spatial cold-start protocol (whole-zone / whole-city holdout, leakage-controlled ablations) to isolate generalization from memorization on a conformal delivery-window model at 90% coverage.
• Quantified epistemic & aleatoric uncertainty (evidential deep learning, MC dropout, a Dirichlet head, and deep ensembles) to drive an abstain-to-SLA fallback pipeline — a 12% relative improvement in >25% window reduction at 90% precision.

NVIDIA Nemotron Reasoning Challenge — teaching a 30B model to reason Fine-tuned a 30B-parameter Nemotron model to solve logic puzzles via supervised fine-tuning with Unsloth. Built an LLM pipeline that classifies every puzzle type and extracts its hidden rule, then used a larger teacher (GPT-OSS-120B, served with vLLM) to generate step-by-step reasoning and distilled it into the smaller model. SFT · PEFT/LoRA · distillation · Unsloth · vLLM

MemEval — does agentic-memory complexity earn its cost? A controlled, reproducible pilot study comparing six LLM-agent memory strategies (no-memory, full-context, BM25, FAISS, Mem0, A-MEM) on the LongMemEval benchmark under one pinned reader, scorer, and budget — measuring accuracy and cost (ingest time, query tokens, latency, store size). Crash-resilient per-system subprocess driver, bootstrap CIs, paired-difference significance testing, and cross-reader replication. agentic memory · evaluation · Ollama · bootstrap stats · reproducibility

Smart Provider Credentialing Analytics Platform — repo A multi-agent system (hand-rolled supervisor: understand → plan → dispatch → synthesize) that reconciles a healthcare provider roster against authoritative sources and answers plain-English questions. The governing rule — the LLM never computes a number — with a deterministic numeric-faithfulness guard, a 3-mode backend (hosted Groq → local Ollama → deterministic fallback), and a 4-suite LLMOps evaluation scorecard gated for CI. multi-agent · LLMOps · faithfulness · DuckDB · Flask · Docker

CAFA-6 Protein Function Prediction — Kaggle Global Rank 19 (Competitions Expert) Placed 19th of thousands worldwide (team). Built a multi-tower model fusing pretrained protein language models (ESM-2, ProtT5) with taxonomy for 3,000-way hierarchical multi-label GO-term prediction, plus a BLAST nearest-neighbor transfer model — ontology-aware label propagation (True Path Rule) and IA-weighted F-max threshold selection throughout. ESM-2 · ProtT5 · multi-label · hierarchical classification · transfer learning

Beetle — hybrid semantic search engine — repo End-to-end retrieval system over AI/ML content: crawling, extraction, weakly-supervised quality classification, and lexical (BM25) + dense (FAISS) + learned-sparse (SPLADE) + ColBERT indexes fused with weighted Reciprocal Rank Fusion and a cross-encoder reranker. Improved NDCG@10 to 0.77 from a 0.65 keyword baseline on BEIR; shipped as a containerized FastAPI service with health probes, property tests, and a DVC pipeline. hybrid retrieval · SPLADE · ColBERT · RRF · FastAPI · DVC · BEIR eval

Epidemic-PINN — recovering hidden transmission rates from case data Recovers an epidemic's time-varying transmission rate β(t) and reproduction number R(t) from reported cases alone, by backpropagating through a differentiable SEIR simulator (torch RK4). On Germany's first wave it tracks an independent OWID estimate to RMSE 0.14 and beats a classical growth-rate baseline by ~10×, with identifiability treated honestly throughout. physics-informed ML · differentiable simulation · inverse problems · PyTorch

B.Tech, Chemical Engineering — Indian Institute of Technology (IIT) Roorkee Expected July 2027 · CGPA 8.39 / 10

Selected self-study: CS229 (Machine Learning, Stanford) · CS224N (NLP with Deep Learning, Stanford) · MIT 6.S191 (Deep Learning)

graph LR
    A[Machine Learning Systems] --> B[LLM Post-training]
    A --> C[Agentic Systems and LLMOps]
    A --> D[Representation Learning]
    A --> E[Information Retrieval]
    A --> F[ML for Science]
    B --> B1[SFT and PEFT]
    B --> B2[Reasoning Distillation]
    C --> C1[Agentic Memory]
    C --> C2[Multi-agent Orchestration]
    C --> C3[Evaluation Harnesses]
    D --> D1[Contrastive Embeddings]
    D --> D2[Protein Language Models]
    E --> E1[Hybrid Retrieval and RRF]
    F --> F1[Physics-informed ML]
    style A fill:#9D4EDD,color:#fff
    style B fill:#7209B7,color:#fff
    style C fill:#7209B7,color:#fff
    style D fill:#7209B7,color:#fff
    style E fill:#7209B7,color:#fff
    style F fill:#7209B7,color:#fff