Software Engineer — Backend Systems & AI/ML Applications

B.Tech CSE (AI & ML), VIT Bhopal · India · Graduating 2027

I build backend systems and AI/ML applications with an emphasis on correctness and reliability — concurrency control, caching, persistence, failure handling, and automated test coverage as first-class design concerns rather than afterthoughts.

My recent work spans backend infrastructure, applied AI, and full-stack systems, with selected applications available as live demos. Each project is documented and backed by an explicit test suite with CI gates.

AI Infrastructure — Inference gateways, request deduplication, circuit breaking, provider failover, hybrid retrieval (BM25 + dense + RRF), token streaming

Backend Engineering — Layered REST architecture, relational modelling, pessimistic locking, keyset pagination, OAuth2/OIDC, RBAC, schema migrations

Testing & Quality — Unit, integration, E2E and concurrency suites; coverage gates in CI; deterministic mocks and offline test harnesses

Applied ML Systems — Retrieval grounding, spaced-repetition scheduling, geospatial classification, evidence-tracked evaluation

Algorithms — C++20, DSA, multi-platform competitive programming

A routed LLM inference gateway built as infrastructure rather than a notebook. Applies TTL+LRU caching, SingleFlight in-flight request deduplication, micro-batching, circuit breaking, and failover across Anthropic, OpenAI-compatible backends (Ollama, vLLM, DeepSeek, Groq), and a deterministic MockProvider so the entire test path runs offline with zero API keys.

Hybrid retrieval fuses BM25 lexical search with dense vectors via Reciprocal Rank Fusion, backed by a RAG evaluation suite (context precision/recall, keyword coverage, faithfulness). Ships SSE token streaming, Prometheus metrics, and an interactive playground.

Measured — 200 concurrent duplicate prompts: 297.7 ms → 45.5 ms wall time via SingleFlight (99.5% deduplication, one provider call).

Runs locally with uvicorn; no public deployment.

40 tests · ruff lint · CI on Python 3.11/3.12 · Docker · 5 ADRs · MIT

Repository · Architecture

A library management platform built to production conventions. Layered REST (Controller → Service → Repository) on Spring Security 6.5 with session auth via BCrypt and opt-in Google OAuth2/OIDC, URL-pattern RBAC across Admin / Librarian / Student, and CSRF protection.

Persistence work includes Flyway-versioned migrations (V1–V5), pessimistic locking for borrow/return race protection, N+1 prevention via EntityGraph, cache eviction on mutation, SQL-level overdue filtering, and keyset pagination for large report queries.

OCI deployment configuration is prepared but no live instance is verified; earlier Railway and Render deployments have been retired.

175 tests / 17 classes · JaCoCo ≥70% gate (90.3% last verified) · Spotless · ArchUnit · 9 concurrency tests · Docker (amd64 + arm64) · 9 engineering docs · MIT

Repository · Deployment guide

A deployed adaptive-learning platform that ingests uploaded material and derives per-topic mastery, confidence, and exponentially decayed retention from recorded attempts.

The distinction that matters: the LLM writes questions and explanations — it never computes mastery, confidence, decay, or review scheduling. Those are deterministic calculations over stored attempts. Quiz generation is synchronous and can take minutes; only indexing and attempt-sync are enqueued to BullMQ, and the API reports honestly when background work was not scheduled.

The shipped retriever is a measured lexical baseline (local-hash-v1), documented as such — a full BM25 implementation exists for comparison but is not on the active path.

Skipped tests are gated behind opt-in flags requiring live infrastructure; they are reported rather than silently passing. CI gates typecheck · lint · build · MIT

Live — athenaeumai.tech · API health · Swagger

Repository

A deterministic pipeline for grounded cross-document fact extraction and epistemic reasoning, built for a VIT engineering internship assignment. Parses multi-page PDFs, extracts structured assertions, and links every fact to a document, page number, and verbatim source sentence before persisting it.

Cross-document fact pairs are classified as CORROBORATES, RECONCILABLE, CONTRADICTS, UNCERTAIN, or UNRELATED — a numeric difference is not automatically a contradiction when estimate vintage or reporting scope explains it. Runs fully offline through an LLMProvider abstraction with a deterministic mock.

Runs locally from a FastAPI server; no public deployment.

300 tests · pytest · hand-verified gold sets (27 extraction / 30 relationship) · precision 1.0000 · F1 0.9200 · relationship accuracy 30/30

Repository

Civic Lens — Geospatial ML for civic complaint triage. KNN spatial-density features, DBSCAN hotspot detection, and XGBoost prioritisation, evaluated on 5,469 real NYC 311 records (95.7% with direct training; 65.3% zero-shot from synthetic data, quantifying cross-city domain shift). The repository documents that the main dataset is synthetic and that high direct-training accuracy should be read cautiously, since status-derived labels partially correlate with the target. React, TypeScript, Firebase, Leaflet. React TypeScript Firebase Python XGBoost · Live

Evidence-Driven Dataset Intelligence — A research artifact, not a production system. Local-first dataset recommendation with a two-layer evidence ledger (observation states and cross-evidence resolution states kept distinct), bounded fingerprinting, hard compatibility gates, multi-objective ranking, and family-aware greedy diversification. Fixed seeds, no network egress during evaluation, and a measured 128 MiB / 20 s envelope. M0–M8 frozen milestones plus an exploratory N = 3 human pilot. 208 tests · M0–M8

AI Smart Retail & Customer Intelligence — Computer vision, NLP, and a hybrid chatbot behind one FastAPI gateway. OpenCV facial encodings stored instead of raw images (privacy by design), MobileNetV2 product classification, and TF-IDF + logistic-regression review sentiment. Live dashboard

Languages — C++20 · Python · Java 21 · TypeScript · JavaScript · SQL

Backend & Infrastructure — Spring Boot · Spring Security · FastAPI · Node.js · Express · REST API design · MySQL · MongoDB · Redis · BullMQ · Flyway · OAuth2/OIDC · RBAC · Docker · GitHub Actions · Linux · Prometheus

AI / ML — Retrieval pipelines (BM25, dense vectors, RRF) · RAG · LLM provider APIs (Anthropic, OpenAI-compatible, Groq) · scikit-learn · XGBoost · OpenCV · PyTorch · SSE streaming · Spaced repetition · RAG evaluation

Testing & Quality — pytest · JUnit / MockMvc · Jest · Vitest · Playwright · JaCoCo · ArchUnit · concurrency testing · deterministic mock providers · coverage gating in CI

I regularly practise data structures and algorithms across competitive-programming platforms to sharpen problem-solving and algorithmic thinking. My solutions, reusable templates and notes are available in the competitive-programming repository, where compilation of every C++ file is validated in CI.

• Model Context Protocol (MCP) and autonomous coding-agent architecture
• Streaming and partial LLM generation to cut end-to-end generation latency
• Sentence-transformer embeddings and cross-encoder reranking as a measured upgrade path over lexical retrieval

Certifications & achievements — TCS CodeVita Season 13 global rank 3,633 · Oracle Certified Foundations Associate in Agentic AI (2026) · NPTEL Introduction to IoT (Elite + Gold, top 5%, 90) · NPTEL Cloud Computing (Elite + Silver, top 2%, 88) · Applied Machine Learning in Python, University of Michigan (Coursera)