I am a B.Tech student specializing in AI & ML who enjoys taking ideas from a model or algorithm and turning them into complete software systems.
My work sits at the intersection of Generative AI, Retrieval-Augmented Generation, machine learning, backend engineering, and data-driven applications. I focus on systems that are structured for evaluation, deployment, observability, and reliability.
- Generative AI & RAG — hybrid retrieval, GraphRAG, embeddings, reranking, evaluation, and LLM applications
- AI Agents — LangGraph workflows, tool calling, planning, memory, routing, and human-in-the-loop execution
- ML Engineering — NLP, forecasting, classification, feature engineering, model evaluation, and deployment
- Backend Engineering — FastAPI, REST APIs, PostgreSQL, Neo4j, authentication, caching, and service design
- MLOps & DevOps — Docker, CI/CD, reproducible benchmarks, monitoring, and production-oriented workflows
| Focus Area | What I am building |
|---|---|
| RAG Systems | Hybrid vector + BM25 + graph retrieval with reranking and evaluation |
| Agentic AI | Stateful workflows with planning, tools, routing, and review loops |
| AI Engineering | FastAPI services around ML and LLM systems |
| System Reliability | Fallbacks, retries, caching, evaluation, and deterministic CI |
Hybrid vector + BM25 + graph retrieval with a corrective fallback layer.
PDF → Parsing → Semantic Chunking → Embeddings → PostgreSQL/pgvector + Neo4j → Query Classification → Vector/BM25/Graph Retrieval → RRF → Cross-Encoder Reranking → LLM Generation
- Hybrid pgvector + BM25 + Neo4j graph retrieval, fused with Reciprocal Rank Fusion and cross-encoder reranking
- Intent-based query classifier routes each question to vector, graph, or hybrid retrieval
- Corrective RAG (v4): a grader agent checks whether the retrieved context actually answers the question, and falls back to live web search instead of guessing
- Measured on a golden set — faithfulness 0.9633, answer relevance 0.4962 → 0.6967 with CRAG, and 0.0440 → 0.6938 on out-of-knowledge-base questions
- Authentication, rate limiting, query caching, session memory, analytics, and evaluation endpoints
- Dockerized services with GitHub Actions CI/CD that runs the benchmark offline (
ENABLE_LIVE_LLM=false) so results stay reproducible
→ Repository: Hybrid-GraphRAG-System · Live: enterprise-hybrid-graphrag.onrender.com
IndicBERTv2 fine-tuned for Hindi NER, with zero-shot transfer to three more languages.
- Fine-tuned on 100k silver-labeled Hindi Naamapadam sentences; 82.74% entity-level F1 on the official gold test set (+8.7 over a BiLSTM-CRF baseline)
- Zero-shot cross-lingual transfer with no target-language training — Bengali 80.82%, Tamil 75.06%, Telugu 85.31% F1
- Bootstrap 95% confidence intervals and error analysis (boundary errors dominate on long 5+ token organization names)
- FastAPI inference service with an interactive entity-highlighting UI, CPU-optimized Docker image, published to Docker Hub
→ Repository: multilingual-indic-ner · Model: Hugging Face
24-hour-ahead household electricity forecasting with calibrated uncertainty.
- LightGBM on 167M smart-meter readings from 4,443 London households, aggregated into a single hourly demand series (19,864 hours, zero gaps)
- 0.0109 MAE — 55.1% below a seasonal-naive baseline (Holt-Winters rejected at +113.3%)
- Split-conformal calibration lifted interval coverage from 64.3% → 82.6% at a nominal 80%, calibrated on validation residuals only
- Error analysis isolated holidays as the dominant failure mode; a controlled holiday-as-Sunday encoding cut holiday MAE 20.6%
- FastAPI recursive-inference service with a test-period replay UI, Docker + GitHub Actions CI/CD, models published to Hugging Face Hub
→ Repository: energy-demand-forecasting · Model: Hugging Face
An experimentation framework where a p-value below 0.05 is never sufficient to ship.
- Two-proportion z-test, chi-square goodness-of-fit for sample-ratio mismatch, and simulation-based Type I error calibration on 294,478 rows
- Bayesian posterior on the same 2×2 table — 94% HDI,
P(B > A), and expected revenue impact — with the README stating plainly that these summaries are not independent confirmations - Verdict DON'T SHIP: the credible interval excludes the +1% worthwhile-lift range,
P(B > A) = 0.1197, and expected revenue is −5,188 - Minimum detectable effect of +2.53% surfaced as a design warning — the experiment was underpowered for the question it was run to answer
- Decision engine with deliberately asymmetric gates: SHIP requires all 4 criteria, DON'T SHIP requires any 1 of 3
→ Repository: ab-testing-experimentation
EfficientNetB0 transfer learning with Grad-CAM explainability and a production FastAPI service.
- 95% accuracy across 4 classes; glioma recall lifted from 65% → 83% via dataset rebalancing plus decision-threshold optimization
- Custom glioma decision threshold, focal loss, and class weights after four iterations on the failure mode
- FastAPI backend with Grad-CAM overlays, batch prediction, confidence thresholds, and request tracing
- Resolved a Keras 2/3 serialization break by rebuilding the architecture in code and loading weights separately
- Docker + GitHub Actions CI/CD publishing to Docker Hub, model hosted on Hugging Face Hub
→ Repository: brain_tumor · Live: brain-tumor-xw48.onrender.com
LangGraph agent that turns a high-level goal into an executable task DAG.
- Planning → execution → review → routing state machine with self-healing retries and dynamic provider fallback across Groq and OpenRouter
- Dependency-aware DAGs of up to 12 tasks, with a clarification flow instead of guessing when the goal is ambiguous
- Lazy-loaded tool registry with 8 tools — file operations, sandboxed shell, web search, and URL fetch
- Session state, human-in-the-loop input, FastAPI service, Docker + CI/CD
→ Repository: Autonomous-Task-Planning-System
CRAG — Corrective RAG · STGCN Traffic Forecasting · Customer Churn API · Crypto Direction Predictor · Heart Disease Prediction · Ticket Categorizer · MLOps Batch Job · RAG Assistant · DSA / LeetCode
I contribute to open-source projects by identifying concrete bugs, improving dataset metadata, adding regression coverage, and extending developer tooling.
Pull Request #2472 · Open
Added a YAML-to-Markdown converter supporting .yaml and .yml files, including nested mappings, lists, tables, multiple YAML documents, anchors/aliases, multiline strings, and edge cases. Added converter registration and comprehensive tests.
Pull Request #14 · Merged
Fixed a FastAPI route-ordering bug where DELETE /api/executions/all was being captured by the parameterized DELETE /api/executions/{execution_id} route. Added regression tests covering both bulk and single-execution deletion.
Pull Request #8 · Merged
Improved secret-leak detection so exact-case and case-insensitive matches are distinguished, with regression tests for both behaviors.
Contributed multiple fixes to awesomedata/apd-core, including:
- Updated the MIT Heart Rate Time Series dataset link to the PhysioNet archive. PR #386 — merged.
- Updated USDA Nutrient Database metadata to reflect FoodData Central. PR #387 — merged.
- Updated the CIMA histological microscopy dataset link to an accessible Kaggle mirror. PR #388 — merged.
- Fixed the Cell Image Library homepage URL and metadata. PR #389 — merged.
→ View Awesome Data contributions
Problem → Data / Documents → Retrieval or ML Pipeline → Model / LLM
→ Evaluation → API / Product → Reliability → Deployment / CI/CD
Retrieval quality · evaluation · latency · failure handling · security · reproducibility · deployment
| Area | Technologies / Concepts |
|---|---|
| Generative AI | LLMs, prompt engineering, RAG, GraphRAG, tool calling |
| Retrieval | pgvector, BM25, graph retrieval, RRF, reranking |
| Agents | LangChain, LangGraph, stateful workflows, memory |
| NLP | TF-IDF, transformers, BERT, classification, semantic similarity |
| Deep Learning | PyTorch, TensorFlow, GNNs, STGCN, transfer learning |
| Machine Learning | XGBoost, Random Forest, Logistic Regression, SVM, clustering |
| MLOps | MLflow, Docker, GitHub Actions, evaluation pipelines |
| Backend | FastAPI, REST APIs, PostgreSQL, Neo4j, authentication |
- Data structures and algorithms
- SQL and data modeling
- System design for AI applications
- Prompt engineering and LLM evaluation
- Experimentation and model analysis
- Writing maintainable, testable Python
I am interested in opportunities and collaborations around AI/ML engineering, Generative AI, RAG, intelligent agents, and backend systems.