Full-Stack, Data & AI Engineer building products, analytics pipelines, and reliable AI systems.
B.Tech CSE - Data Science & Analytics @ IIIT Nagpur - 2027
Data Analytics & Engineering - Full-Stack Products - AI / ML & Agent Systems
- Data analytics & engineering - SQL transformations, CDC and ETL pipelines, analytics models, Power BI/Metabase dashboards, statistical analysis, and forecasting that connect data to business decisions.
- Full-stack products - React/TypeScript interfaces, Node.js/Express and FastAPI services, databases, authentication, billing, webhooks, and third-party integrations.
- AI / ML systems - Stateful agents, evidence-grounded retrieval, local model inference, vision-language fine-tuning, human review, evaluation, and failure handling.
I care as much about what happens when the model is wrong as when it is right.
| Project | What it demonstrates | Tech |
|---|---|---|
| CortexAI - AI Developer Platform & Integration Workspace | Full-stack AI workspace with a React client, Express API gateway, auth/chat/agent/billing services, MongoDB and Redis persistence, LangGraph routing, RAG, multimodal workflows, and generated artifacts. | Node.js, Express, React, MongoDB, Redis, LangGraph, Qdrant |
| North Star - Governed Expense Operations Platform | Expense operations with deterministic policy, 13 durable n8n workflows, advisory AI, human approvals, SHA-256 provenance, notifications and SSE. Local release reports 182 PostgreSQL tests, plus a Metabase bootstrap with 36 questions and five dashboards. | Python, FastAPI, React, TypeScript, PostgreSQL, n8n, MCP |
| SifraAI - Deployed Voice AI Product | Deployed MERN-style voice assistant product with Firebase auth, cookie/JWT sessions, MongoDB, Razorpay billing, Gemini, an embeddable JavaScript widget, and third-party website integration. | React, Node.js, Express, MongoDB, Firebase, Razorpay, Gemini |
| AI Platform Reliability Copilot | FastAPI incident clustering, risk scoring and runbook retrieval with Redis, OpenTelemetry, Prometheus/Grafana, Slack and a six-stage CI pipeline. Public API health reports v2.0.0 with a mock LLM provider; the demonstration uses synthetic telemetry. | Python, FastAPI, Redis, OpenTelemetry, Prometheus, Grafana, Helm |
| RevenueGuard AI - Explainable Payment Recovery | Event-driven payment-recovery platform with webhook ingestion, ML triage, gateway-health intelligence, Redis-backed workers, PostgreSQL, LangGraph decisions, deterministic policies, and human approval for high-value actions. | Python, FastAPI, LangGraph, PostgreSQL, Redis, Razorpay, React |
| StyleHub - Xeno CRM Backend | Campaign manager with SQL-grounded customer segmentation, a hybrid deterministic/LangGraph agent, human launch approval and lifecycle journeys; paired with a separate channel simulator for asynchronous delivery and receipt callbacks. | FastAPI, Streamlit, SQLite, SQLAlchemy, LangGraph |
| Project | What it demonstrates | Tech |
|---|---|---|
| Retail Lakehouse Engineering | Replay-safe PostgreSQL CDC through Airflow and Spark/Delta into dbt silver/gold models; 42,796 local CDC events, 38/38 dbt build, SCD snapshots, metric-grain checks, and a PostgreSQL EXPLAIN/OLTP-to-OLAP lab. | PostgreSQL, Airflow, PySpark, Delta Lake, dbt, AWS S3 |
| RideIQ 2.0 - NYC Demand Forecasting & Fleet Intelligence | Analytics over 9.55M taxi trips with leakage-safe zone-hour forecasting, route clustering, a dashboard/API, and constrained allocation simulation. Random forest test WAPE: 18.2% versus 21.3% for the strongest temporal baseline. | Python, Pandas, Scikit-learn, Streamlit, FastAPI |
| Retail Analytics & Data Pipeline System | Azure-ready ingestion and PySpark transformation into a five-dimension/one-fact star schema; 10,201 orders, 18 DAX measures, a three-page Power BI report and a two-page Azure DirectQuery report. Uses simulated/public data. | Azure Data Factory, Azure SQL, PySpark, Power BI, SQL, DAX |
| Vendor Performance Analysis | Python/SQL ingestion, vendor-brand aggregation, Power BI and statistical analysis of concentration, profitability and inventory. The analyzed population shows 65.69% purchase concentration in the top 10 vendors and $2.71M estimated unsold stock; observational findings. | Python, SQLAlchemy, SQLite, SQL, Power BI, SciPy |
| BrandPulse Analytics | Full-stack e-commerce analytics with Pandas ETL, PostgreSQL, a Flask REST API, Redis-cached KPIs, React charts, CSV export and ETL run history. Docker Compose supports local operation; Azure deployment is a proposed extension. | React, Flask, PostgreSQL, Redis, Pandas, Docker |
| Manufacturing Quality & Equipment Anomaly Intelligence | Synthetic manufacturing telemetry transformed through PostgreSQL/dbt into quality, anomaly, downtime and OEE-style analytics served by FastAPI and Metabase; 14 dbt models and 31 data tests verified locally. | Python, PostgreSQL, dbt, Scikit-learn, FastAPI, Metabase |
| Project | What it demonstrates |
|---|---|
| SignalBrief - Competitive Intelligence | Deployed research-to-review workflow: Zapier RSS ingress, CrewAI evidence collection, AutoGen critique, human approval and Slack delivery. Neon preserves source hashes, execution records and a transactional outbox; 49 regression tests reported on SQLite and PostgreSQL. Public approved-run workspace. |
| TripBandhu - Stateful Agentic Travel Research | v1.1.1: five LangGraph specialists, typed provider evidence, PostgreSQL checkpoints, provider reliability controls, interrupt/resume review and session-bound server-rendered PDFs. 188 regression tests and 49/49 FAST benchmark cases reported. |
| MyManah Journal Intelligence | Local-model journal analysis and cited PDF Q&A with FastAPI/React, Ollama, Hugging Face classifiers, SQLite, Chroma and hybrid retrieval. Current policy revision reports 96 deterministic Windows tests; actual-model, offline and backup/restore checks are documented. Engineering prototype, not clinically validated. |
| Internal RFP Analyst | Cyclic RAG with citation verification and 145 tests on main. The separate Neo4j GraphRAG edition adds vector/graph/hybrid modes and reports 641 passed, 5 skipped; its public walkthrough used transparent vector fallback, and comparative retrieval superiority is not claimed. |
| Financial Document Intelligence | SEC EDGAR ingestion, section-aware chunks, dense/BM25 retrieval, grounded citations, abstention and evaluation tools; 79 tests reported. Full indexing runs locally; the hosted UI is a lightweight preview. |
| EnergyShield | Three coordinated workflows connect geopolitical signal interpretation, deterministic disruption scenarios and weighted supplier ranking, with persisted reasoning and decision support. Hackathon prototype using mock/API-ready signals. |
AI is part of my normal build loop, but I do not treat "the agent says it works" as completion. I use coding agents for repo exploration, implementation, tests, repetitive changes, and debugging, then verify behavior with tests, CI, traces, logs, and the running system. The architectural and product decisions stay mine.
| Area | Tools and Technologies |
|---|---|
| Backend & APIs | Node.js, Express, FastAPI, REST APIs, service boundaries, authentication, sessions, webhooks, billing |
| Full-stack | React, TypeScript, JavaScript, MongoDB, PostgreSQL, Redis, SQL, SQLAlchemy |
| Agents & GenAI | LangGraph, LangChain, MCP, RAG, tool calling, structured outputs, HITL, guardrails, provenance, agent evaluation |
| Reliability & Infrastructure | pytest, GitHub Actions, Docker, LangSmith, observability, SSE, n8n, failure handling |
| Integrations | AWS S3, Qdrant, Tavily, Metabase, Resend, Slack, Firebase, Razorpay |
| Data Engineering | PostgreSQL CDC, Airflow, PySpark, Delta Lake, dbt, replay-safe ingestion |
| Analytics & BI | Pandas, NumPy, SQL, Power BI, DAX, Metabase, statistical tests, forecasting, data quality, dimensional modeling |
| Local & Multimodal ML | PyTorch, Transformers, PEFT/LoRA, SmolVLM, Ollama, Chroma, Whisper, pyannote |
| Project | What it demonstrates | Tech |
|---|---|---|
| ReceiptKIE-VLM - V2 High-Resolution Receipt Extraction | Continued LoRA adaptation of SmolVLM at 1536 px; 99.2% valid JSON, 63.3% macro field exact match and 18.3% complete-record exact match on 246 previously unseen SROIE test receipts. V1 remains a reproducible baseline. | PyTorch, Transformers, PEFT/LoRA, SmolVLM, SROIE |
| SpeakerScan | Speech annotation combining pyannote diarization, wav2vec2 emotion and Whisper language ID with sequential model loading, resumable stage checkpoints and atomic JSON/RTTM exports plus a CSV manifest. | Python, PyTorch, pyannote, wav2vec2, Whisper, Streamlit |
IIIT Nagpur. B.Tech CSE - Data Science & Analytics. Graduating 2027. Interested in roles where software, data and AI meet: full-stack engineering, data analytics/engineering, and AI/ML systems. Available immediately for a 6-month internship across these areas.
Project summaries checked against current repository documentation on 9 October 2026. Local tests, offline evaluations and prototypes are identified separately from deployed products.
Contribution data refreshes daily from GitHub. Profile views are approximate image requests, not unique visitors; the badge is cached and may lag.
I am open to Full-Stack, Data Analytics / Engineering, and AI / ML internships and collaborations involving useful products, trustworthy data and reliable systems.
- Email: tg304429@gmail.com
- Portfolio: tushar-portfolio-taupe.vercel.app
- LinkedIn: linkedin.com/in/tushar-ghosh-a3355124a
- GitHub: github.com/tusharg007