Independent AI & Systems Engineering · Selected Builds

Engineering intelligent systems with evidence, not just impressive demos.

Hari Nandan K builds across applied AI, analytics, retrieval evaluation, defensive security and physical-systems software. The work spans autonomous analysis pipelines, scientific evaluation harnesses, environmental data systems, digital signal processing and embedded design.

Testable engineering
Data provenance
Explicit security boundaries
ENGINEERING INDEX // HN-2026
SELECTED BUILDS
Hari Nandan K
Hari Nandan K
AI · Data · Security · Physical Systems
46
AI Scientist Tests
10 suites · repo test matrix
6
RAG Architectures
Comparable evaluation harness
8
Fraud Models
Synthetic graph benchmark
3
Engineering Domains
AI/data · security · physical systems
> EVIDENCE: Repository test matrices & benchmark harnesses
> BOUNDARIES: Synthetic data, simulations & prototypes labelled

Systems built to be examined

Three projects that best represent my current direction: autonomous analysis, scientific evaluation and graph-based risk modelling. Each card links to its source and states what the available evidence actually supports.

Autonomous AI Data Scientist analyst workspace
Research build · 46 tests reported passing
AUTONOMOUS AI / ANALYTICS ENGINEERING

Autonomous AI Data Scientist

A controlled analysis workflow that turns business questions into evidence-linked SQL, statistical investigation and reviewable reports. The system is designed to refuse unsupported questions rather than invent missing data.

Engineering breakdown Problem · architecture · evidence · limits
Problem
Text-to-SQL systems can select the wrong grain, create join fanout, mutate data or produce plausible explanations unsupported by the warehouse.
System design
A 21-state bounded workflow with a canonical semantic layer, SQL AST validation, hypothesis testing, an independent critic and human review checkpoints.
Architecture
Python, FastAPI, DuckDB and a React + TypeScript analyst workspace. SHA-256 dataset references bind generated claims to the data used for analysis.
Repository evidence
The committed TEST_MATRIX.md records 46 passing tests across 10 suites, covering warehouse generation, semantic validation, SQL safety, statistics, causal checks, evidence verification and API flows.
Scope limit
The bundled warehouse and scenarios are synthetic. Passing tests verify implemented code paths; they do not establish accuracy on external customer data or production business outcomes.
PythonFastAPIDuckDBReactTypeScriptSQL ASTBayesian StatisticsSHA-256 Evidence
RAG evaluation and benchmarking workspace
Evaluation framework · 34 tests reported passing
RETRIEVAL SYSTEMS / EMPIRICAL EVALUATION

RAG Evaluation Platform

A reproducible harness for comparing retrieval-augmented generation architectures under a shared evaluation protocol, with retrieval, grounding, robustness, latency and cost measurements.

Engineering breakdown Problem · architecture · evidence · limits
Problem
RAG demos often report anecdotal quality without comparable datasets, retrieval metrics, adversarial cases or operational cost measurements.
Architectures
Naive, hybrid BM25+dense with RRF, reranked, query-rewriting/HyDE, GraphRAG and agentic RAG, executed through a common harness.
Evaluation
Precision@k, Recall@k, MRR, NDCG@k, faithfulness, citation precision/recall, adversarial robustness, latency percentiles and token/cost models, with paired statistical comparisons.
Repository evidence
The committed TEST_MATRIX.md records 34 passing tests across 9 suites, covering ingestion, retrieval stores, six architectures, metrics, robustness, statistics, storage and API execution.
Scope limit
The included challenge sets are synthetic. Results from this harness are dataset- and configuration-dependent, not a universal ranking of RAG architectures or proof of customer-corpus performance.
PythonFastAPIReactBM25RRFHyDEGraphRAGStatistical Testing
Real-time graph fraud detection system
Synthetic benchmark · 8-model progression
GRAPH ML / FINANCIAL RISK RESEARCH

FraudShield AI — Real-Time Graph Fraud Platform

A streaming fraud-research platform that models relationships among users, cards, devices, IPs, merchants and addresses, then compares tabular, anomaly-detection and graph-learning approaches.

Engineering breakdown Problem · architecture · evidence · limits
Problem
Isolated transaction features can miss coordinated patterns that emerge across linked entities and change over time.
Architecture
A temporal heterogeneous graph, streaming feature windows, point-in-time state, drift monitoring and a FastAPI + React investigation interface.
Model comparison
Eight model families from logistic/XGBoost baselines and anomaly detectors to GCN, GraphSAGE, GAT and Temporal GNN, compared with leakage-aware temporal evaluation.
Repository evidence
The README reports 24/24 tests passing and approximately 12.8 ms P99 typical latency for its stated local workload. Treat that latency as a repository-reported synthetic benchmark, not a financial-network SLA.
Scope limit
The simulator and injected fraud patterns are synthetic. This is not connected to a bank or live payment processor and has not established production fraud-detection performance.
PythonPyTorchPyTorch GeometricFastAPIGraph Neural NetworksSSETemporal Evaluation

From technical question to reviewable system

I take on bounded engineering problems where architecture, implementation and validation matter. Scope and acceptance criteria are agreed before delivery.

Applied AI & Retrieval Systems

Build AI-assisted workflows with explicit data boundaries, retrieval evaluation and human review where the output is consequential.

  • Document ingestion and retrieval pipelines
  • Tool-using and multi-step agent workflows
  • RAG evaluation, regression tests and traceability
  • Structured outputs, API integration and review flows

Data & Analytics Engineering

Turn data sources into clear, inspectable analytical workflows instead of ungrounded narratives or one-off dashboards.

  • SQL-backed analytics and semantic data models
  • FastAPI services and analytical data stores
  • Statistical evaluation and experiment reporting
  • Interactive dashboards and evidence-linked reports

Defensive Security Tooling

Develop defensive utilities and controlled research environments with clear authorization boundaries and explainable outputs.

  • Static file triage and integrity workflows
  • Access control, audit logging and evidence chains
  • Lab SIEM pipelines and detection prototypes
  • Threat-data visualization and analyst workflows

Embedded & Applied Systems

Bridge software, signal processing and hardware-oriented engineering through simulation, firmware references and documented design work.

  • ESP32-oriented embedded software and tooling
  • DSP workflows and measurement interfaces
  • Telemetry ingestion and digital-twin prototypes
  • KiCad design documentation and engineering reviews

A Disciplined 5-Step Process

From initial scope to a documented delivery, with acceptance criteria and known constraints made explicit.

Explore the delivery process 5 stages
01

Discover

We align on your exact product objectives, target users, technical constraints, and definition of success through a focused scoping session.

02

Architect

I design the system architecture, database models, AI/LLM integration strategy, and security boundaries before writing code.

03

Build

Rapid, modular development in iterative sprints with clean, type-safe code, automated test coverage, and transparent asynchronous progress demos.

04

Validate

Run project-specific tests, evaluate outputs against the selected datasets, inspect failure modes and record unresolved gaps.

05

Deploy & Support

Prepare the deployment or handover appropriate to the agreed scope, document operation and known limitations, and define support boundaries.

Curiosity across software and physical systems

A practical engineering approach: understand the system, make assumptions explicit, build a testable implementation and show where the evidence stops.

I am Hari Nandan K, an independent builder working across applied AI, data systems, defensive security, environmental intelligence, digital signal processing and embedded engineering. My repositories range from autonomous analysis and retrieval evaluation to geospatial research, audio DSP and hardware design.

I care about the engineering beneath the interface: data contracts, state machines, SQL safety, evaluation methodology, provenance, access control, latency budgets and how a system behaves when its assumptions fail.

I also work on education and community tooling through the HOPE initiative. The goal is to make technical learning more practical through interactive documents, coding tools, project workflows and AI-assisted learning experiences.

I distinguish a passing test suite from field validation, synthetic data from measured data, and a digital hardware design from a fabricated board. That distinction is part of the engineering, not a footnote.

Evaluation before claims

Use test matrices, benchmark protocols and explicit datasets to show what has actually been exercised.

Traceable outputs

Preserve data lineage, hashes, source context and the reasoning behind system decisions.

Bounded security work

Prefer defensive tooling, explicit authorization and isolated lab environments.

Software meets hardware

Explore embedded firmware, DSP, telemetry systems and computer-aided hardware design.

Hari Nandan K

Build carefully. Measure honestly. Explain the limits.

Tools used in the work

A repository-grounded stack grouped by engineering domain. Inclusion means a tool appears in selected project code or documentation; depth varies by project.

Explore technologies and methods 5 engineering domains

AI, ML & Retrieval

Agents · model evaluation · inference
  • Multi-step orchestration and bounded state machines
  • RAG, BM25, reciprocal-rank fusion and GraphRAG
  • PyTorch, graph learning and anomaly detection
  • Bayesian reasoning and statistical evaluation
PythonPyTorchscikit-learnOllamaBM25GraphRAGHyDE

Backend & Data Systems

APIs · analytical storage · event flow
  • REST APIs, streaming endpoints and service integration
  • SQL validation, semantic models and data contracts
  • Evidence stores, hashing and experiment persistence
  • Telemetry ingestion and event-time processing
FastAPIFlaskNode.jsExpressDuckDBPostgreSQLSQLite

Web & Interactive Systems

Interfaces · visualization · connectivity
  • Responsive application interfaces and analyst workspaces
  • Interactive 3D visualization and scientific views
  • WebRTC peer connectivity and real-time streams
  • Progressive web application workflows
TypeScriptJavaScriptReactViteTailwind CSSThree.jsWebRTCSSE

Defensive Security

Integrity · access · detection research
  • RBAC, signed tokens and authorization boundaries
  • SHA-256 integrity and tamper-evident audit records
  • Static file inspection and safe quarantine workflows
  • Synthetic SIEM telemetry and Sigma-subset detections
HMAC-SHA256SHA-256YARASigmaRBACMITRE ATT&CK

Embedded, DSP & Hardware

Firmware references · signal processing · EDA
  • ESP32-S3-oriented firmware reference and C/C++ development
  • LMS, NLMS, FIR and real-time audio analysis
  • KiCad schematics, PCB constraints and design reviews
  • SystemVerilog and simulation-led engineering studies
C / C++ESP-IDFNumPySciPyPortAudioKiCadSystemVerilog

Experiments, prototypes and deeper systems

A curated view of the broader repository portfolio. Project status, dataset scope and validation limits are stated explicitly; synthetic benchmarks and digital designs are not presented as field deployments.

Open the R&D archive 17 projects · searchable
Showing 17 of 17
ALEX environmental observation platform Research platform · provider status varies
ENVIRONMENTAL INTELLIGENCE / REMOTE SENSING

ALEX — NASA-MODIS / VIIRS Intelligence

An Earth-observation research platform focused on MODIS and VIIRS fire-observation harmonization, spatial and temporal evidence, provenance, incident workflows and role-based access.

TypeScriptMODIS / VIIRSGeospatialProvenanceRBAC
Scope & limitations Read before interpreting results
Evidence boundary
The repository documents a mix of operational, credential-dependent, fail-closed and simulation-mode integrations. External provider availability and live credentials are not implied by this portfolio.
AFTERLIGHT planetary archaeology visualization Research submission · 6 tests documented
3D SCIENCE / PLANETARY HERITAGE

AFTERLIGHT — Planetary Archaeology

An interactive Three.js platform for exploring spacecraft heritage across the Solar System, with lunar and Martian globes, rover traverse paths, 3D hardware inspection and source-linked scientific claims.

Three.jsWebGLScientific Data3D VisualizationProvenance
Scope & limitations Read before interpreting results
Evidence boundary
The repository's test report documents six unit tests and browser checks. Scientific coordinates and archival records still depend on the cited source data and stated uncertainty.
V.O.I.D.E. industrial telemetry concept MVP · synthetic telemetry
INDUSTRIAL TELEMETRY / DIGITAL TWIN

V.O.I.D.E. — Telemetry Intelligence

An industrial telemetry workbench with event-time ingestion, schema checks, deterministic replay, per-asset twin state, health indicators and a statistical baseline for chiller-like systems.

PythonFastAPICSV IngestionDigital TwinStatistical Baselines
Scope & limitations Read before interpreting results
Evidence boundary
The V.O.I.D.E. 2.0 README explicitly uses reproducible synthetic telemetry. Risk and remaining-useful-life estimates are heuristic; no real-plant validation or safety certification is claimed.
SENTINEL-X SIEM dashboard concept Lab-only · synthetic events
CYBER DEFENCE / SIEM RESEARCH

SENTINEL-X — Streaming SIEM

A local lab SIEM that normalizes endpoint and network-style events, applies deterministic YAML and Sigma-subset detections, correlates alerts into incidents and exposes hunt and metrics interfaces.

PythonFastAPISigma SubsetMITRE ATT&CKSQLitePrometheus
Scope & limitations Read before interpreting results
Evidence boundary
The README states that seeded adversary-like records are inert synthetic data and instructs users not to connect the lab to customer, public or third-party networks.
AEGIS-X security operations console Research console · emulated telemetry
SECURITY OPERATIONS / AGENTIC SYSTEMS

AEGIS-X — Autonomous Security Operations Console

A security-operations research console combining multi-tier signal handling, specialist analysis agents, Bayesian evidence fusion, incident context and cryptographic audit records.

Node.jsAgent WorkflowsBayesian FusionSHA-256Security Analytics
Scope & limitations Read before interpreting results
Evidence boundary
The public README describes a research console and emulated telemetry. Do not interpret agent counts, risk scores or simulated incidents as live SOC deployment evidence.
Q-SAFE local malware triage tool Early-stage defensive tool
LOCAL SECURITY / FILE TRIAGE

Q-SAFE — Local-First Malware Triage

A defensive file-inspection utility using static format checks, entropy and packing indicators, suspicious-string heuristics, SHA-256 reporting and an explicit reversible quarantine flow.

PythonStatic AnalysisSHA-256YARAClamAVcapa
Scope & limitations Read before interpreting results
Evidence boundary
The README explicitly says this is not certified antivirus or production EDR. Samples are not executed, and heuristic findings are triage leads—not proof of malware. Detection has no published independent held-out benchmark.
ADNR adaptive audio workstation Desktop engineering app · 33 tests badge
DIGITAL SIGNAL PROCESSING / AUDIO

ADNR — Adaptive Noise Reduction Workstation

An offline desktop DSP environment with dual-channel audio processing, LMS/NLMS adaptive filters, FIR processing, live spectrum and convergence views, experiment logging and an ESP32-S3 C/ESP-IDF reference.

PythonPySide6NumPySciPyPortAudioESP32-S3 Reference
Scope & limitations Read before interpreting results
Evidence boundary
The repository includes a desktop application and embedded C reference implementation. That is not evidence of a fabricated or deployed ESP32 hardware product.
Chameleon peer-to-peer communication concept Prototype · peer connectivity required
PRIVACY / PEER-TO-PEER COMMUNICATION

Chameleon — P2P Chat, Voice & Video

A browser-native peer-to-peer communication experiment using WebRTC for direct text, media and call sessions, with AES-GCM encryption and an installable web-app interface.

TypeScriptWebRTCAES-GCMPWAPeer-to-Peer
Scope & limitations Read before interpreting results
Evidence boundary
Connectivity depends on browser, signaling and network traversal conditions. The repository's security design should not be read as a completed independent security audit or a guarantee of zero metadata.
J.A.R.V.I.S. desktop assistant Prototype · Linux desktop
LOCAL AI / DESKTOP AUTOMATION

J.A.R.V.I.S. — Linux Desktop Assistant

A Linux-oriented assistant experiment combining local Ollama inference, a Flask and Socket.IO interface, system telemetry, Three.js visualization and multi-step terminal workflows.

PythonOllamaFlaskSocket.IOThree.jsLinux
Scope & limitations Read before interpreting results
Evidence boundary
This is a personal assistant prototype. Local model quality, available tools and terminal permissions determine what it can safely execute.
F.R.I.D.A.Y. browser voice assistant Experimental browser prototype · no backend
BROWSER AI / VOICE INTERFACE EXPERIMENT

F.R.I.D.A.Y. — Single-File Browser Assistant

A zero-build browser assistant experiment packaged in a single HTML file, with a voice interface, installable PWA behavior, offline-capable client state and a responsive interface.

HTMLJavaScriptWeb Speech APISpeechSynthesisPWAlocalStorage
Scope & limitations Read before interpreting results
Evidence boundary
Speech recognition availability varies by browser, locale and permissions. Speech synthesis uses browser-provided voices, and offline behavior depends on cached resources. This is a browser-side experiment, not a secure operating-system assistant.
ORION-X FPGA accelerator PCB design Digital hardware design · DRC findings open
FPGA / HIGH-SPEED PCB ENGINEERING

ORION-X R1 — FPGA PCIe Accelerator

A KiCad design package for a Kintex-7 FPGA accelerator concept with PCIe Gen2 x4, DDR3L memory, a 12-layer stackup, signal- and power-integrity documents, schematics and manufacturing outputs.

KiCadFPGAPCIe Gen2 x4DDR3LSI / PI12-Layer PCB
Scope & limitations Read before interpreting results
Evidence boundary
The latest checked DRC report, generated 17 November 2026, lists 69 DRC violations, including dangling tracks and vias. It is not DRC-clean or fabrication-ready; closure and independent engineering review remain necessary.
MOSAIC Sigma multi-physics design concept Simulation-led design · not fabricated
MULTI-PHYSICS / EXPERIMENTAL COMPUTING

MOSAIC-Σ — Autonomous Matter Intelligence

A digital research design for a modular instrument concept combining microwave metasurface interrogation, ultrasonic and impedance measurements, analog reservoir computation and Bayesian experiment selection.

KiCadSystemVerilogPythonRF ConceptsEISBayesian Design
Scope & limitations Read before interpreting results
Evidence boundary
The repository explicitly states that physical fabrication is excluded. Reported results are based on numerical or circuit simulation, not measurements from a fabricated instrument.
LexiQuery AI legal document assistant v0.1.0 · active development
DOCUMENT AI / DOMAIN-SPECIFIC RAG

LexiQuery AI — Legal Document Assistant

An early legal-document Q&A application built around PDF ingestion, parent-child retrieval and citation-oriented answers, with an emphasis on preserving clause context.

PythonRAGPDF IngestionSemantic RetrievalCitations
Scope & limitations Read before interpreting results
Evidence boundary
This is a software prototype, not legal advice or a substitute for qualified counsel. Citation-oriented output does not guarantee that every answer is correct or complete.
WhatsApp automation suite Application build · consent safeguards
MESSAGING / WORKFLOW AUTOMATION

WhatsApp Automation Suite

A Node.js and React messaging workflow application with multi-device connectivity, campaign composition, scheduled follow-ups, contact management, response inbox, opt-out handling and rate controls.

Node.jsExpressReactViteBaileysSQLite
Scope & limitations Read before interpreting results
Evidence boundary
The public README describes Baileys-based connectivity, not an official Meta Cloud API integration. Deployments must follow platform policies, consent requirements and applicable messaging laws.
HOPE education platform Education platform · active codebase
EDTECH / LEARNING SYSTEMS

HOPE Edu-Hub — Academic Super-Platform

An education-oriented application codebase with AI tutor workflows, document and spreadsheet tools, presentation and report generation, project design, learning roadmaps and community features.

ReactViteSupabaseAI TutorHope DocsLearning Tools
Scope & limitations Read before interpreting results
Evidence boundary
This is an evolving platform codebase. Capabilities described in project documentation should be checked against the current deployment and configured providers before being treated as generally available.
Autonomous agentic BI workspace Prototype · validation pending
BUSINESS INTELLIGENCE / ANALYTICS

Autonomous Agentic BI

A 16-stage business analytics workflow with semantic grounding, SQL validation, statistical decomposition, counterargument review, evidence-linked claims and a deliberate refusal path when available data cannot support a question.

PythonFastAPIDuckDBReactSQL ASTStatisticsEvidence Store
Scope & limitations Read before interpreting results
Evidence boundary
The repository's committed TEST_MATRIX.md marks its test suites PENDING. This is a design and implementation prototype, not a validated analytics product.
Personalized email generator agent Prototype · agent workflow
AGENTIC AUTOMATION / SALES RESEARCH

Personalized Email Generator

A Next.js and FastAPI application that researches publicly available prospect context and uses an LLM workflow to prepare tailored sales-email drafts.

PythonFastAPINext.jsOpenRouterResearch Agent
Scope & limitations Read before interpreting results
Evidence boundary
Automated prospect research may be incomplete or inaccurate. Outputs require human verification, and outreach must follow privacy, anti-spam and platform rules.

What has been tested — and what has not

A public portfolio should let reviewers distinguish automated tests, synthetic benchmarks, simulations and real-world validation.

Automated test matrices

Tests attached to source

The AI Data Scientist repository records 46 passing tests across 10 suites; the RAG Evaluation repository records 34 passing tests across 9 suites. These are repository-level test results, not a production uptime or accuracy guarantee.

Data provenance

Synthetic data is labelled

The RAG and fraud projects use synthetic challenge data; V.O.I.D.E. 2.0 uses synthetic telemetry and separate ground-truth files. No real customer or plant performance is inferred from those datasets.

Security boundaries

Defensive tools have stated limits

Q-SAFE is early-stage static triage, not certified antivirus or EDR. SENTINEL-X is a synthetic lab SIEM and is not intended for connection to third-party networks.

Hardware readiness

Digital design is not fabrication

ORION-X's latest checked DRC report lists 69 findings dated 17 November 2026. MOSAIC-Σ is simulation-led and explicitly not fabricated. Neither is described here as a validated physical product.

Let's Build Something Exceptional

Have an ambitious product idea, need an intelligent automation tool, or want to augment your engineering team? Submit an inquiry below.

Direct Partnership

I work with founders, startups and product teams to scope and build applied AI, analytics, automation and software systems with clear deliverables and technical handover.

Availability

Available for scoped engineering engagements

Location

Kozhikode, Kerala, India (Remote Worldwide)

Project Inquiry Form

Direct enquiry · Scope and requirements reviewed before commitment
×