Somitra Singh
Kushwah.

Full-stack agentic RAG · FastAPI + React + PostgreSQL

A multi-user RAG platform built around a LangGraph tool-calling agent that autonomously decides between document retrieval, live Gmail actions, and memory recall per query — genuine agentic reasoning instead of static retrieve-then-generate.

• Hybrid dense + sparse retrieval: pgvector HNSW + BM25 (ParadeDB), fused via Reciprocal Rank Fusion — all from a single Postgres instance.
• Persistent long-term memory built from scratch: LLM-driven write tool, two-layer semantic deduplication, automatic cross-session recall per user.
• Content-aware ingestion for PDFs, images, and 10+ formats (Docling, Unstructured) with LLM metadata enrichment and table-preserving chunking.
• Multi-user security: Supabase auth, per-thread authorization, Fernet-encrypted per-user OAuth credentials for Gmail — concurrency-safe under simultaneous load.

Self-hosted agentic deployment · Linux VPS

Deployed production agent instances on a VPS, exposed through Telegram and Discord bots, and used them as daily drivers for code, research, and content workflows.

• Configured MCP server integrations, tool permissions, and sandboxed execution.
• Built async webhook + long-poll bridge in Python with per-chat session state and rate limiting.
• Hardened with systemd, log rotation, reverse proxy, and secrets management — codified real failure-mode fixes into restart-safe configs.

LangGraph multi-node agent · GPT-4o

A conversational booking agent built on LangGraph's stateful graph with conditional edge routing — a hands-on take on the ReAct pattern.

• 4 custom LangChain tools with Pydantic-validated schemas, backed by parameterized SQL against PostgreSQL (Neon).
• Persistent conversation memory via LangGraph's PostgresSaver, plus per-user long-term memory for preferences.
• Streamlit chat UI with Google OAuth 2.0 (PKCE) and user-scoped data isolation.

Config-driven daily agent · GitHub Actions

A fully automated agent that runs daily on GitHub Actions to fetch AI/ML job listings, filter by recency, deduplicate, and persist results — with LLM scoring on top.

• Pluggable provider abstraction (Adzuna, JSearch/RapidAPI) normalizing heterogeneous API responses into a unified schema.
• OpenAI JSON-mode scoring of each posting's relevance 0–100 with rationale, plus tailored resume generation with guardrails against fabricating experience.
• CI pipeline with fail-fast secret validation, artifact upload, conditional auto-commit, concurrency control, and pip caching.

End-to-end ML pipeline · AWS SageMaker

A classification model predicting phone price ranges, with the full train/deploy/invoke loop running on SageMaker — S3 for data, IAM for least-privilege access.

• Custom training script + SageMaker estimator for cloud training and evaluation.
• Scalable prediction endpoint with a clean upload → train → deploy → invoke workflow.

End-to-end MLOps · Terraform-provisioned AWS

A phishing-URL detection pipeline covering ingestion, schema validation, KNN imputation, training, and serving — every stage modular and independently testable.

• Compared six classifiers with hyperparameter search; auto-selected the best model (F1 0.976, precision 0.962, recall 0.99) with MLflow + DagsHub tracking.
• FastAPI inference service (/train, /predict) containerized with Docker; GitHub Actions CI/CD builds, pushes to ECR, and rolls the ECS service on every push to main.
• Entire AWS stack as code with Terraform — ECS Fargate, ECR with scan-on-push, encrypted S3 artifacts, CloudWatch, and least-privilege IAM roles.