I'm Anupam Kumar, an AI Engineer and Computer Science graduate from IIITM Gwalior ('25). I build production-oriented AI systems across RAG, agentic workflows, retrieval, computer vision, and AI backends.

My focus is on making AI systems reliable beyond the demo — from retrieval and grounding to evaluation, backend architecture, and deployment.

Currently open to AI Engineering, ML/AI Infrastructure, and Platform Engineering roles.

Always interested in working with people who think deeply, build seriously, and aren't afraid to challenge how something is supposed to be done.

Languages

Python C++ JavaScript

AI / LLM

LangChain PyTorch HuggingFace Groq RAG Agentic AI

Retrieval & Vector Search

FAISS Pinecone ChromaDB BM25

Backend & Data

FastAPI PostgreSQL Redis MongoDB Celery

DevOps, Cloud & Observability

Docker AWS Nginx Git Prometheus Grafana MLflow

Frontend

React Tailwind CSS Streamlit

A citation-graph-aware retrieval system for computer vision research — dense, lexical, and citation-graph signals fused into a single retrieval layer, with grounding measured rather than assumed.

Most RAG systems treat retrieval as one similarity search. AetherCV treats it as three independent evidence sources — semantic similarity, exact terminology via BM25, and citation relationships between papers — merged through a router that decides which signal to trust for a given question, then verifies the answer is actually supported before it's returned.

• Multi-signal retrieval (dense + BM25 + citation-graph) reaches 0.94 context recall, measured against a held-out evaluation set rather than reported from a single run.
• A 7-layer Redis caching architecture (exact, semantic, retrieval, decomposition, intent) pushes the cache hit rate above 84%, holding p99 latency near 4.9 seconds on CPU-only infrastructure — no GPU in the serving path.
• A multi-signal semantic router (domain centroid, retrieval-support probe, entity-shape check) catches out-of-domain questions before generation, keeping every answered response grounded in retrieved evidence with zero false-positive escapes.
• Fully instrumented for MLOps — MLflow experiment tracking, Prometheus/Grafana monitoring, containerized deployment — with zero evaluation failures across all benchmark runs.

Python FastAPI FAISS BM25 Redis PostgreSQL Prometheus MLflow Docker

Live Demo · GitHub

A single natural-language query — "electronics stores in Bengaluru" — becomes validated, AI-scored leads with drafted outreach, end to end, in one API call.

At its core is a six-stage identity resolution engine solving a problem most lead-gen tools quietly get wrong: confirming a website actually belongs to this business, in this location — not a same-named competitor three cities over, or an unrelated directory listing. Evidence, features, and five independent verifiers feed a confidence-propagation model, so every match ships with an explainable score instead of a guess.

• Benchmarked on 113 real-world queries across 20+ Indian cities: 88.8% correct website resolution, 87.4% validation success, 96.4% cross-provider agreement, 100% query-parse reliability.
• A six-tier escalating scraper (static fetch → TLS-fingerprint impersonation → headless rendering → multi-page crawl) feeds a three-tier enrichment waterfall and four LangGraph-orchestrated agents — Company Intelligence, Decision, Review, Messaging — each backed by a deterministic fallback, so no stage depends on an LLM to function.
• Grounding is independently measured, not assumed: an evaluation harness runs the full production pipeline end to end, cross-checking every AI-claimed signal against scraped evidence before it reaches a lead record. The Decision agent can only downgrade a lead's priority, never inflate it; outreach falls back to strict templates rather than let a model invent facts.

FastAPI SQLAlchemy LangGraph LangChain Groq (Llama 3.3 70B) Playwright curl_cffi Next.js 14 TypeScript TanStack Query Zustand Prometheus / Grafana

Live Demo · GitHub

AyurGenix — Agentic RAG over 10,000+ pages of Sanskrit manuscripts. Custom Char-CNN OCR for Sanskrit script, Pinecone-indexed sub-second retrieval, cross-encoder reranking, and citation-grounded answers via a LLaMA-3 conversational layer. FastAPI PyTorch Pinecone LangChain — GitHub

TalentForge AI — Autonomous job-application pipeline: LinkedIn scraping → LLM resume scoring (70% semantic, 30% keyword) → Playwright-based Easy Apply, gated by a strict state machine and a 4-tier LLM fallback chain for unattended, reliable operation. Python Playwright Groq LangChain Streamlit — GitHub

Indian Institute of Information Technology and Management, Gwalior B.Tech in Computer Science — 2021 to 2025