A personal air-quality health-risk correlator, built India-first: aligns lagged exposure windows (t-0/6h/24h/72h) against a personal symptom log, trains a personal RandomForest model with a population-model fallback and time-based (not random-split) evaluation, and answers questions through a grounded WHO/CPCB RAG advisory with an explicit refusal path when the corpus doesn't support an answer. The dashboard leads with the CPCB NAQI and GRAP-stage awareness for Delhi-NCR, and every exposure window carries a stated data_coverage_pct instead of pretending irregular real-world sensor data is clean.

FastAPI Next.js PostgreSQL + pgvector scikit-learn

A 7-layer autonomous defense pipeline: eBPF kernel probes for syscall-level telemetry, deep packet inspection on encrypted traffic without decrypting it, a 4-model ML ensemble (XGBoost, Random Forest, Autoencoder, Isolation Forest), an on-device LLM (Llama 3.2) that turns raw detections into MITRE ATT&CK-mapped explanations, and a Neo4j attack-graph engine with risk propagation — all streamed live to a React dashboard over Kafka/WebSockets.

• 99.16% accuracy · 0.9995 AUC-ROC across 1.1M+ samples
• Built to solve real SOC pain: signature-based IDS misses zero-days and encrypted C2 traffic, and analysts drown in >90% false-positive rates
• Every alert is fully attributed: ML-scored, LLM-explained, and graphed in attack-path context

Python XGBoost PyTorch eBPF Kafka Neo4j FastAPI React

A full-stack disaster intelligence platform running on live USGS seismic and weather data: an ML risk-prediction engine, an NLP layer reading social media for panic signals, a per-user Individual Vulnerability Score, graph-based evacuation-route optimization (NetworkX), satellite hotspot analysis, a shelter/resource allocation optimizer, and a responder AI copilot — with multi-channel real-time alert dispatch over WebSockets and offline PWA support for low-connectivity disaster zones.

FastAPI Next.js NetworkX HuggingFace WebSockets

Two layers that work together: a live operations console visualizing services, network topology, security posture, and ML models — and real Terraform/Docker/Kubernetes IaC that a user can run against an actual AWS account to produce the infrastructure the console visualizes, including a genuinely working SageMaker fraud-detection endpoint and CI/CD in GitHub Actions and Jenkins. Runs standalone as a demo, or for real against live cloud.

Terraform Docker Kubernetes AWS SageMaker Node.js

A locally-trained ML classifier layered with LLM-reasoned verdicts (Groq), OCR-capable document/batch ingestion, and an AI summarizer chatbot, presented through a custom "verification desk" editorial UI. Trained on a 13,343-row balanced dataset across 13 topic buckets — and shipped with a diagnosed root cause for an accuracy drop from a benchmark-looking 95% down to a real-world 86%, traced to narrow-dataset overfitting and corrected rather than hidden.

FastAPI SQLAlchemy React Groq OCR

Four independent services — frontend, API gateway, a Python AI service, and a background worker — talking over HTTP and a shared Postgres/Redis backend, built on one rule: the fast path never blocks on the slow path. Document upload returns instantly; extraction, OCR, and LLM reasoning run asynchronously on a durable BullMQ queue that survives crashes. The AI service only trusts requests carrying an internal shared token, so an accidental port exposure still isn't an open door.

Node.js Python BullMQ Postgres Redis

A multilingual (13 Indian languages) legal-guidance platform: describe a legal problem in plain language and get back your rights — each tied to a named Act and Section, never a vague generalisation — concrete next steps, a procedural timeline, a ready-to-fill document (complaint letter, legal notice, FIR draft), and a follow-up chat scoped to the situation. Six domains are covered (consumer, tenancy, employment, family, criminal, cyber), each backed by a transparent, inspectable keyword-scoring intent classifier instead of a black-box model — so a "why was I routed here" question always has a real answer.

• Urgency detection flags situations implying immediate danger and surfaces emergency contacts (NALSA, Police, Women's/Child helplines) ahead of regular guidance
• Auto-generates fillable legal documents with live preview and client-side PDF export
• Deliberately backend-free — intent classification, document generation, and session state all run client-side or via a single serverless route — with the Groq-backed chat degrading gracefully to an offline rule-based engine when unconfigured

Next.js 16 TypeScript Tailwind CSS v4 Groq

Turns Agmarknet's inconsistent mandi-price data into something a farmer can act on: the real price nearby with an honest read on how stale it is, the best market to sell at after transport cost, a 7-day price forecast, a government-scheme eligibility check, and a voice-first way to ask any of it. The engineering centre is two pipelines — fuzzy-matched commodity/variety normalization that never force-matches or drops a record, and multi-source time alignment that flags every imputed value with a days_stale count instead of hiding the gap.

Next.js FastAPI PostgreSQL + pgvector scikit-learn

Government water-quality bulletins are published as inconsistent PDFs — different columns, different units, some scanned images instead of text. Taproot runs a tiered extraction pipeline (pdfplumber/PyMuPDF, falling back to OCR) behind a strict Pydantic validation gate: the LLM never writes to the database directly, and anything it can't confidently structure goes to a human review queue instead of being dropped or guessed. Ships with real anomaly detection (rolling z-score, IQR, IsolationForest), a cited RAG chat over the bulletins, and an explicitly-labelled turbidity-from-photo heuristic rather than a claimed lab measurement.

Next.js FastAPI PostgreSQL + pgvector

A question goes in; five specialized agents — planner, searcher, summarizer, fact-checker, writer — collaborate to produce a sourced report where every claim in the prose carries a visible, hoverable verdict: Verified, Plausible, Unverified, or Contradicted. Built entirely on free-tier infrastructure (Groq LLM, DuckDuckGo search, no paid API), and designed to degrade gracefully rather than crash when the free tier gets rate-limited.

FastAPI React Groq DuckDuckGo Search

Literature Review, Hypothesis Generation, Experiment Design, and Results Analysis agents run a genuine cyclical loop (LangGraph), each cycle's findings feeding the next — literature-grounded before a hypothesis is allowed, pre-registered before an experiment runs, and scored by real scipy statistics rather than an LLM guessing a p-value. Every ResultsAnalysis is explicitly flagged as single-cycle and pending replication, and the loop pauses for human approval by default — a research accelerator, not an autonomous scientist.

FastAPI LangGraph React scipy Groq

Ingest → Clean → Analyze → Visualize → Report, as five stateless agents behind a REST API and a deliberately zero-build vanilla JS frontend. Cleaning proposes a dry-run diff; nothing touches the data until a human approves it. Every statistical finding is backed by a real scipy test — t-test, ANOVA, Pearson, regression — never a heuristic dressed up as significance.

Flask pandas scipy Vanilla JS

A folder-tree academic resource browser — notes, papers, slides — with recursive global search, open browsing/download with no forced signup, and a single audited admin account (JWT + bcrypt) for uploads and management. Rebuilt twice: from a shared-token prototype to a real per-account auth system with a full login audit trail, without ever touching the neo-brutalist frontend identity that made v1 worth keeping.

Node.js Express React JWT Tailwind

Resume-to-role fit scored by a deterministic, weighted algorithm — 50% skill coverage, 30% experience relevance, 20% project quality — layered with live job-market data, ATS compatibility checks, and resume optimization suggestions. Shipped with a documented post-launch audit that caught and fixed 7 real issues, rather than a README that pretends the first build was clean.

FastAPI Next.js TypeScript Groq

An India medical-cost estimator with government health-scheme eligibility checks and empanelled-hospital surfacing. The one non-negotiable design rule: no cost figure or eligibility result is ever generated by an LLM — every number traces back to a structured JSON data source, and every approximation is explicitly labeled as one instead of being silently substituted.

FastAPI Pydantic Next.js TypeScript

Clusters public complaints via HDBSCAN embeddings, ranks severity, and routes each cluster to the responsible authority. Classification is deliberately deterministic and auditable — LLM usage is scoped only to summarization — because a government user needs to be able to defend a system-assisted decision, not just trust a black box.

FastAPI PostgreSQL pgvector HDBSCAN

A multi-agent finance system — budgeting, debt, tax, and retirement agents — where every calculation is 100% deterministic and independent of LLM availability, including an employer-match-gap analyzer built on real marginal tax-bracket data. The LLM explains; it never computes.

FastAPI React Groq SQLite

Merges Open Library and Project Gutenberg into a single catalog with content-based recommendations, reading-level scoring, and reading-streak tracking — built as a single deployable Next.js app on purpose, trading a heavier stack for something that's actually simple to run and reason about.

Next.js TypeScript SQLite