Projects & Open Source Repositories
Production AI pipelines, localized RAG engines, multi-agent frameworks, computer vision classifiers, and developer infrastructure tools. Detailed architecture breakdowns for each system.
Risk Minor
PRODUCTIONMulti-agent risk intelligence platform built for regulatory compliance workflows at Hansraj Ventures. Combines dense and sparse retrieval with a fine-tuned LLM, full evaluation tracing, and input/output safety guardrails in a single coherent pipeline.
// Architecture
- Hybrid retrieval: BM25 keyword search + Milvus dense vector search running in parallel; results merged via reciprocal rank fusion (top_k=12) before passing to the LLM context window.
- LLM layer: LoRA-fine-tuned Llama 3.1 8B served via FastAPI with CUDA inference; guardrails enforce output format, hallucination checks, and regulatory policy compliance before any response is returned.
- Observability: Full RAGAS eval harness (faithfulness, answer relevancy, context recall) + Langfuse distributed tracing across all pipeline nodes — every inference logged with latency, token counts, and eval scores.
- Infra: Async FastAPI microservices containerized with Docker CI/CD; 40% latency reduction, 99%+ uptime in production at Hansraj Ventures.
Banking Compliance RAG
OPEN-SOURCEEnd-to-end financial compliance system combining classical machine learning with retrieval-augmented generation to answer natural language regulatory queries over banking policy documents in real time.
// Architecture
- ML layer: Multivariate regression pipeline trained on loan-default and risk-scoring datasets; integrated directly into response generation to provide quantitative compliance context alongside retrieved text.
- Retrieval: FAISS dense vector index + Qdrant persistent vector DB for hybrid retrieval across RBI and banking regulation corpora; BM25 reranking applied before final context assembly.
- Validation: Citation verification module cross-checks every LLM output claim against the source document passages; rule-based validators enforce domain-specific banking policy constraints.
- Storage: SQLite for document metadata, conversation history, and audit logs; LangChain orchestrates the full retrieval-generation chain.
AstraRoute Navigation Engine
OPEN-SOURCEFull-stack, disaster-resilient navigation platform engineered for high-risk climate corridors. Dynamically calculates safe perimeter detours around real-time climate hazards and breaking news bounding boxes.
// Architecture
- Disaster Scraping: Background workers monitor news RSS feeds using Scrapling and BeautifulSoup to auto-extract disaster locations into spatial hazard boxes.
- Perimeter Routing: Evaluates OSRM polyline intersections against MongoDB hazard coordinates to dynamically compute cardinal (N, S, E, W) detour vectors.
- Live Weather & Crowdsourcing: Queries Open-Meteo forecast API across route waypoints while accepting authenticated community hazard reports on Leaflet.js.
AI Blur & Defect Detector
OPEN-SOURCEAssessment-ready computer vision pipeline and REST API that automatically evaluates image quality metrics, detects structural blur, exposure anomalies, and noise levels.
// Architecture
- Quality Analytics: Computes OpenCV Variance of Laplacian for blur detection alongside grayscale standard deviation differences for noise estimation.
- Scoring & Persistence: Combines multi-issue metric deductions into an interpretable 0–100 quality score, logging full evaluation history to SQLite.
- Deployment: Asynchronous FastAPI backend paired with a dynamic JavaScript UI, containerized using Docker Compose for instant deployment.
Fact Check Multi-Agent (VeriCore)
OPEN-SOURCEFake-news and misinformation detection engine that queries three locally-running LLMs in parallel and aggregates their verdicts through a consensus scoring layer, dramatically reducing single-model hallucination and bias.
// Architecture
- Model ensemble: Qwen2:0.5B, Phi-3:latest, and LLaMA 3.1:latest queried simultaneously via Ollama; each returns a structured verdict (TRUE / FALSE / UNVERIFIED) with confidence score and reasoning chain.
- Consensus engine: Weighted majority voting across the three verdicts; disagreements trigger a fourth "tiebreaker" call with a larger model context window.
- Live context: Optional DuckDuckGo web search module injects real-time news snippets as additional context before the consensus round — addresses recency gaps in local model training data.
- Evaluation: RAGAS faithfulness metrics validate each model's reasoning against its cited context; disagreement deltas tracked per model to identify consistent bias patterns.
Intelligent Proxy Gateway (IPG)
OPEN-SOURCEFour-stage LLM request optimization middleware deployed as both a Flask/FastAPI backend and a Chrome Manifest V3 extension — routes queries intelligently between local on-device models and cloud LLMs based on task classification and cost.
// Pipeline Stages
- Stage 1 — Prompt Reformulation: Rewrites incoming queries for improved clarity, specificity, and instruction alignment before routing begins.
- Stage 2 — Security Filtering: Scans for prompt injection patterns, jailbreak attempts, and PII leakage using rule-based and embedding-similarity checks; blocks or sanitizes before forwarding.
- Stage 3 — Task Triage: Classifies request complexity (simple/medium/complex) and routes to the appropriate model: local Ollama models (Mistral, Phi-3) for lightweight tasks; cloud LLM fallback (OpenAI/Anthropic) for complex reasoning.
- Stage 4 — Token Optimization: Compresses context window before final LLM call via extractive summarization, reducing token cost by up to 35% on long-context requests.
RAG · Airbus FCOM
RESEARCHUltra-high-precision retrieval-augmented question-answering system grounded in Airbus Flight Crew Operating Manual (FCOM) documentation — a domain where factual errors are operationally unacceptable.
// Architecture
- Document processing: FCOM PDFs parsed with structure-aware chunking preserving section hierarchy, chapter numbering, and callout blocks; custom metadata tagging enables procedure-level retrieval.
- Retrieval: Dense vector index with re-ranking optimized for precision over recall — every retrieved chunk must be from the correct section and version of the manual.
- Faithfulness guardrail: Every generated answer traced back to source passage citations; any answer with insufficient grounding returns a confidence warning rather than a fabricated response.
Local Co-pilot
OPEN-SOURCEOn-device AI coding assistant extension for VS Code, built on Ollama for fully offline, zero data-leakage developer workflows with real-time code completion and context-aware suggestions.
// Architecture
- Context engine: Sliding window context manager maintains relevant code context (open file, cursor position, recent edits, imports) for accurate completions without exceeding model context limits.
- Local inference: Routes completions to Ollama-served models (Codestral, DeepSeek-Coder, Phi-3) — no data leaves the machine; works fully offline.
- VS Code API integration: Custom inline completion provider with debounced triggering; separate command palette integration for longer generation tasks like docstring generation and test scaffolding.
Image Enhancer
OPEN-SOURCEComputer vision pipeline combining super-resolution and intelligent denoising for automated image quality improvement — built as companion infrastructure for specialized vision models with degraded input data.
// Architecture
- Super-resolution: ESRGAN-based upscaling pipeline achieving 4× resolution enhancement with perceptual loss training for natural detail preservation over pixel-accuracy MSE.
- Denoising: Adaptive bilateral + DnCNN hybrid approach; noise level estimation module selects denoising strength dynamically per-image rather than using a fixed filter.
- Pipeline integration: Exposes a REST API endpoint for batch processing and integrates as a preprocessing stage for downstream vision classifiers — designed for VRAM-constrained environments.
Potato Disease Classifier
OPEN-SOURCECNN-based agricultural leaf disease detection system identifying early blight, late blight, and healthy potato plants from field images — designed for high-throughput inference in resource-limited agricultural environments.
// Architecture
- Model: Custom CNN architecture trained on the Plant Village dataset; data augmentation pipeline (rotation, flipping, color jitter, Gaussian noise) improves generalization to real-world field photography conditions.
- Optimization: Post-training quantization to INT8 reduces model size by 75% while maintaining accuracy within 2% of the full-precision baseline; enables deployment on edge devices and mobile inference.
- Inference throughput: Batch inference pipeline with TFServing for production deployment; ~120 images/sec throughput on a single consumer GPU.
Civic Connect Engine
OPEN-SOURCESmart India Hackathon (Internal Round Winner) project — CNN classifier for detecting urban civic hazards (potholes, garbage, damaged infrastructure) from crowdsourced citizen photos, integrated with GCP microservices for automated complaint routing.
// Architecture
- Classifier: Transfer-learning on MobileNetV2 fine-tuned on a custom civic hazard dataset; achieved 91% validation accuracy across 8 hazard categories with a balanced precision-recall tradeoff.
- Crowdsourced pipeline: Mobile photo uploads routed through GCP Cloud Storage → Cloud Run inference container → automated complaint ticket creation in the civic management system based on predicted hazard type and severity score.
- Infra: Fully containerized with Docker; GCP Cloud Run autoscaling handles variable citizen submission volumes; image metadata (GPS, timestamp) attached to every complaint for geographic routing.
NetScan Pro
OPEN-SOURCEHigh-performance network reconnaissance utility for discovering active hosts, enumerating open ports, fingerprinting service banners, and building a structured inventory of network topology — built as a developer security research tool.
// Architecture
- Host discovery: Multi-threaded ICMP ping sweep with configurable concurrency and timeout parameters; CIDR range parsing for subnet-level scans.
- Port scanning: SYN, TCP Connect, and UDP scan modes with configurable port ranges; banner grabbing for common service identification (HTTP, SSH, FTP, SMTP, etc.).
- Output: Structured JSON and CSV export of scan results with service version fingerprints; designed for integration into broader security pipeline workflows.