Projects

Built local. Verified against real failure cases.

Each system pairs a transparent scoring engine with a documented bug it actually caught — not a happy-path demo. 37 fully-local platforms in total; 23 featured below.

CyberShield AI

Local AI cybersecurity threat detection platform

+
PythonFastAPIscikit-learnOllamaSQLAlchemyStreamlit
builtIntrusion detection, malware/phishing classification, and automated incident response via a local IsolationForest anomaly model plus a transparent weighted-rule engine.
signatureEvery score returns its exact contributing factors, with Ollama-generated explanations and deterministic fallback playbooks.

ContractMind AI

Local contract review & compliance platform

+
OllamaChromaDBTesseract OCRsentence-transformers
builtOCR ingestion, clause extraction, compliance checking, and contract comparison behind a weighted-rule risk engine.
signatureRisk engine returns exact triggered rules per clause; a RAG legal Q&A pipeline that only answers from retrieved clauses and always cites sources.

RetailVision AI

Smart retail analytics platform

+
YOLOv8OpenCVPostgreSQL
builtYOLOv8-based person detection with a centroid tracker for footfall and occupancy, plus a classical-CV fill-ratio engine for shelf inventory.
signatureUpgrades gracefully to a custom-trained model once labeled data becomes available.

SmartFactory AI

Industrial predictive maintenance platform

+
XGBoostscikit-learnPandas
builtUnsupervised IsolationForest anomaly detector — no labeled failure data required — plus an XGBoost Remaining Useful Life predictor with a trend-based fallback.
signatureEvery health score traces back to its exact contributing sensor features.

GeoVision AI

Satellite image intelligence platform

+
GDALrasterioPostGIS
builtExplainable rule-based NDVI/NDWI/NDBI classifier for land use/cover — zero training data required.
signatureA magnitude-based change detector that correctly distinguishes vegetation loss from flooding.

EvalBench AI

Local LLM evaluation & benchmarking platform

+
OllamaStreamlitPlotly
bugMissing reference context was silently scored as grounded during hallucination detection.
fixVerified end-to-end that ungrounded claims now fail the check rather than pass by default.

Local AI Gateway

Local LLM infrastructure & observability platform

+
FastAPIhttpxChart.jsOllama
builtSingle-API gateway unifying multiple local LLMs behind unified chat/embeddings/rerank endpoints, with automatic failover and response caching.
signatureLive dashboard for p50/p95 latency, cache hit rate, and per-model usage, with an in-browser completion playground.

CloudOps AI

Local AIOps & infrastructure intelligence

+
scikit-learnOllamaSQLite
bugPost-incident data was silently widening the model's definition of "normal," causing false-positive flapping.
fixFixed-snapshot calibration and incident cooldown — false positives dropped near zero while 15/15 real incidents still detected.

PromptOps AI

Local prompt versioning, evaluation & A/B testing

+
SQLitehttpxChart.js
builtTreats prompts as source code: immutable version control with diffs, pluggable evaluators, A/B testing, tolerance-banded regression testing.
resultIterated a classification prompt from 67% to 100% eval pass rate while cutting latency 10x; surfaced a run-to-run variance bug motivating multi-sample comparison.

FeatureStore AI

Local feature management & serving platform

+
FastAPISQLite
builtVersioned schemas, offline/online store separation, point-in-time correct retrieval, training/serving skew detection.
verifiedA query as-of a past date returns exactly what was known then; skew detection isolated a deliberately-injected 37% distribution shift without flagging unshifted features.

LLMSec AI

Local LLM security & red-teaming platform

+
SQLiteChart.jsOllama
bugA narrow, high-signal jailbreak leak marker was being outvoted by accumulated-weight thresholding, misclassifying real jailbreak compliance as merely partial.
fixVerified against a regression suite confirming genuine refusals still classify correctly.

Federated Learning AI

Local privacy-preserving collaborative training

+
scikit-learnNumPyChart.js
builtReal FedAvg training across simulated clients, non-IID partitioning, differential privacy with a tracked accumulating budget, additive secure aggregation.
verifiedSecure aggregation matches plain averaging exactly (masks cancel); non-IID partitioning surfaces genuine per-client variance instead of smoothing it away.

SmartCity AI

Urban analytics & traffic intelligence

+
YOLOv8XGBoostOpenCV
builtVehicle counting, congestion forecasting, incident detection, smart parking, with explainable signal-timing optimization via iterative water-filling.
bugFixed a real defect where minimum-green floors silently exceeded the requested cycle length.

SearchMind AI

Enterprise hybrid search platform

+
ChromaDBrank-bm25sentence-transformers
builtBM25 + vector search via reciprocal rank fusion, access-role-aware retrieval, grounded RAG Q&A.
verifiedAccess control fails closed on malformed metadata, and filters before ranking — not after.

RoboMind AI

Robotics vision & navigation platform

+
OpenCVYOLOv8A*
builtA* path planning, reactive obstacle avoidance, confidence-weighted sensor fusion — testable without ROS2, Isaac Sim, or hardware.
verifiedGenuine edge cases: out-of-bounds cells, obstacle-occupied vs. free-but-enclosed unreachable goals — not just the happy path.

ManufacturingQA AI

Visual quality inspection platform

+
OpenCVscikit-learnPostgreSQL
builtClassical-CV defect detection, golden-sample anomaly detection, and dimensional inspection.
bugFixed a contour-detection bug flagging part outlines as false defects while missing real scratches inside — verified via clean-vs-defective test images.

TravelMind AI

Intelligent travel planning platform

+
OllamaChromaDBsentence-transformers
builtRAG destination research, weather-aware scheduling, budget planning, hotel matching.
bugFixed a scheduling bug where activity exhaustion and genuine bad weather produced the same misleading reasoning text.

InsuranceIQ AI

Insurance claims intelligence platform

+
Tesseract OCROpenCVscikit-learn
builtOCR field extraction, policy validation, and damage assessment behind a transparent weighted-rule fraud scorer combined with unsupervised anomaly detection.
signatureEvery risk score traces back to its exact contributing rule or feature.

EduMentor AI

Personalized learning and tutoring platform

+
OllamaChromaDBsentence-transformers
builtMastery tracking and weighted-adaptive quizzes with a ChromaDB + Ollama RAG pipeline that answers only from retrieved course material.
signatureAlways cites sources, falling back to raw chunks if the LLM is unavailable — never a silent, ungrounded answer.

EnergyOpt AI

Smart energy forecasting and optimization platform

+
XGBoostscikit-learnPandas
built24h demand forecasting, peak load prediction, and carbon tracking using recursive XGBoost prediction.
signatureAn IsolationForest anomaly detector with an explicit deviation-from-baseline feature to catch anomalies in the current reading itself.

ResearchMind AI

Research paper assistant

+
OllamaChromaDBsentence-transformers
builtSemantic search, citation-aware RAG Q&A, and grounded summarization answering strictly from retrieved content.
signatureA literature-review generator that deduplicates by paper.

ClimateVision AI

Climate and weather intelligence platform

+
XGBoostscikit-learnPandas
builtXGBoost forecasting, extreme-weather anomaly detection, and EPA-standard air quality monitoring.
bugFixed an explainability bug where contributing features were ranked by raw magnitude instead of standardized z-score — verified against an injected heatwave.

BioVision AI

Medical imaging engineering demo (non-diagnostic)

+
OpenCVPostgreSQL
builtAn explicitly non-diagnostic medical-image processing demo — candidate-region flagging, Otsu segmentation, explainability heatmaps.
signatureTreats explainability and the disclaimer as one requirement: every flagged region's bounding box and z-score IS the full output.

Plus 20+ additional platforms across document intelligence, finance, agent orchestration, and voice AI — see below, or browse everything on GitHub.

Developer, Document & Vision Tooling

CodeMind AI (repo-aware code Q&A, LangGraph + AST safety layers), Recommend AI (hybrid recommendation with drift detection), DocFlow AI (local OCR-to-RAG document intelligence), AI Video Intelligence System (real-time detection, tracking, scene summarization).

Finance & Operations Intelligence

FraudShield AI (real-time fraud scoring, XGBoost), SupplyChain AI (demand forecasting, route optimization via OR-Tools), MarketMind AI (stock forecasting, LSTM/XGBoost with sentiment).

Retrieval-Augmented Systems

FinSight AI (financial RAG over 10-K filings), HireMind AI (resume-job matching with explainable scoring), MedInsight AI (clinical decision support, tested on 50+ real documents).

Developer & Agent Tooling

AutonoMind AI (multi-agent orchestration), AgentOps AI (agent observability, tracing, cost analytics), DataForge AI (distributed ETL/ELT pipelines).

Monitoring & Voice AI

MLOps Platform (drift detection across 45+ trained models, KS-test + PSI), VoiceFlow AI (real-time voice assistant with conversational memory, Whisper).

Research

Crop Prediction & Recommendation System — Random Forest, 99% accuracy across 22 crop classes and 7 environmental parameters. Presented at 14th ICCET 2026 (Paper ID: ICCET260383).