7+ years bridging safety-critical automotive systems engineering (EPS, Steer-by-Wire, ISO 26262, ASPICE) with production AI automation. Built and deployed enterprise AI platforms at Bosch Global Software Technologies, delivering $86,719 USD in verified cost avoidance and 2,760+ engineering hours saved across 10–15 European and North American OEM vehicle programs. Previously digitalized the full test & validation lifecycle at ZF Rane.

"The tool nobody talks about — but every requirements engineer desperately needs."

Problem	Importing customer specification PDFs into IBM DOORS required up to 15 days of manual reformatting per project — error-prone, repetitive, and blocking every downstream activity
Solution	Enterprise document transformation pipeline built on Azure Document Intelligence, with a dual-path extractor alongside PaddleOCR-VL → interactive React canvas block editor → concurrent export to three DOORS-compatible formats (RTF with OLE, HTML, ReqIF XML)
Key engineering	• ReqIF 1.2 compliance — full OMG spec with multi-tool targeting (DOORS Classic 9.x, DOORS Next/ELM, Polarion, Elektrobit) · • Two-tier automatic ReqID discovery — structural fingerprinting + frequency scoring, with meta-pattern regex fallback · • Three-tier heading inference — model → regex numbering → bbox height heuristic · • OLE COM pooling manager — process recycling to eliminate GDI handle leaks · • OCR result caching — immutable/mutable JSON split avoids re-running GPU inference on revisits · • Language filter — auto-discards non-English blocks from mixed content
Impact	15 days → ~3 days (80% cut) · 6,000+ pages processed · $30,163 USD cost avoidance · Currently in testing for worldwide enterprise release across Bosch
Tech	Azure Document Intelligence PaddleOCR-VL FastAPI React/Vite/TypeScript ReqIF 1.2 Word COM (OLE) Zustand
Status	· 🏛️ Enterprise Architecture Showcase

"Stop re-inventing requirements that already exist in your legacy projects."

Problem	Engineers manually compared incoming customer requirements against 10+ legacy project specifications — up to 15 days per review cycle, zero consistency
Solution	ChromaDB-backed semantic search adopted across 10–15 European and North American OEM vehicle programs (Steer-by-Wire, EPS, Braking), built around a dual-path token saver: exact-string bypass at $0 LLM cost + embedding search for the rest — cutting API costs 40%
Key engineering	• 5-Gate self-improving prompt compiler — (1) statistical error pattern analysis on aggregated feedback, (2) LLM-designed prompt patch, (3) 4-check automated validation (stat backing, contradiction detection, shadow test on holdout set, confidence threshold), (4) human review, (5) canary deployment with deterministic 10% session routing and auto-promote/rollback · • Level 1 learning — feedback recall skips the LLM for previously-reviewed pairs (cosine ≥ 0.97) · • Per-user skill learning — extracts matching preferences from corrections, persists in SQLite, injects top-10 rules into the prompt
Impact	15 days → 5 days (65% cut) · $56,556 USD cost avoidance · 1,800+ engineering hours saved
Tech	ChromaDB NVIDIA NIM FAISS Streamlit OpenAI SDK Tenacity
Status	Deployed internally · 🔗 Public POC

"Catch what a rushed manual review misses — before it reaches the vehicle."

Problem	Verifying safety-critical requirement attributes against ISO 26262 and SOTIF relied entirely on manual expert review — slow, and inconsistent under deadline pressure
Solution	LangGraph multi-agent compliance verification graph combining specialized agents — Safety Standard Parser, ASIL Decomposition Auditor, Requirement Verifiability Checker, Safety Critic — with evaluator-optimizer loops
Impact	>90% recall on safety-critical attributes, validated against certified functional safety experts · 60% faster review cycle
Tech	LangGraph ISO 26262 ISO 21448 (SOTIF) Multi-Agent Evaluator-Optimizer
Status	· Internal enterprise specification honored with Bosch Engineering Excellence Award (July 2024)

"Ask your engineering documents anything — text, tables, drawings, audio, video."

Problem	Engineers had no unified way to query across heterogeneous document types (PDFs, drawings, audio meeting notes, video recordings) to verify functional safety requirements
Solution	Corrective RAG (CRAG) pipeline orchestrated by LangGraph with 5 agents — Router → Retriever → Grader → Rewriter → Generator. Self-correcting loops rewrite and re-retrieve (HyDE, step-back, decomposition) whenever retrieved context is irrelevant
Key engineering	• Multimodal ingestion — PDF/DOCX/PPTX (Docling OCR + table extraction), audio (Groq Whisper-large-v3), video (ffmpeg keyframe extraction + VLM captioning), images (PyMuPDF + VLM) · • Hybrid vector search — Qdrant dense + BM25 sparse with RRF fusion, followed by an NVIDIA reranker · • Content safety — NVIDIA NemoGuard 8B pre-screens all queries · • FastMCP server — exposes indexing and search as standard Model Context Protocol tools via SSE · • 298 automated tests across the pipeline
Tech	LangGraph Qdrant NVIDIA NIM (Llama 3.2 VLM) Supabase FastMCP Next.js Docling
Status	· 🔗 Live Demo · 📦 Public Repository

"What if I change the torsion bar diameter from 9mm to 10mm? — answered in seconds, not days."

Problem	Impact analysis of design changes required consulting multiple disconnected specification documents, with knowledge scattered across teams
Solution	Hybrid agentic pipeline: entity extraction → Neo4j knowledge-graph multi-tool query (11 tools) → sufficiency check → scoped ChromaDB vector search (4 tools) → Claude Sonnet synthesis with extended thinking. 12 granular SSE event types stream reasoning steps live
Key engineering	• Cytoscape.js renders Neo4j traversal paths as interactive graph visualizations · • Plotly.js projects vector search results as 2D PCA cluster scatter plots · • Versioned prompt system — 8 purpose-specific prompts · • Drawing OCR — Gemini API extracts structured data from engineering drawings with sufficiency scoring · • Dual-run tolerance stack engine — flags radial clearance violations flipping PASS to FAIL
Tech	Next.js 14 FastAPI (SSE) Neo4j ChromaDB Claude Sonnet Gemini Cytoscape.js
Status	· 🏛️ Enterprise Architecture Showcase

"From paper-based chaos to real-time digital test operations."

Problem	The entire test lifecycle — request, scheduling, execution, reporting — ran on paper and spreadsheets
Solution	End-to-end system: Request → Scheduling → Execution Tracking → Automated Report Generation
Role	Project Lead & Process Architect — defined business logic and system architecture, managed the external dev team
Impact	80% paperless · real-time tracking across 20+ hydraulic test rigs and DAQ stations
Tech	Process Design .NET Data Management
Status	Deployed in production

"Not just detection — a full driving decision system for Indian mixed traffic."

• Layer 1 — Detection: YOLOv8m trained on the IDD dataset with progressive resolution (640→960→1280px) and heavy weather augmentation (rain, fog, sun flare, motion blur)
• Layer 2 — Inference engine: Detection → IoU Tracking → Monocular Distance Estimation → Behavior Classification → 6-Level Safety Decision Hierarchy → Decision Smoothing
• Key metrics: mAP@50-95: 0.420 · Precision: 0.773 · F1: 0.655 · INT8 quantized: 3× size reduction, 2× FPS gain
• 🥇 1st Prize — Autonomous Driving & Edge Computer Vision Hackathon
• 🔗 Full repository

"Same accuracy, a fraction of the wait."

• Quantized a local vision-language model using llama.cpp and GGUF
• 4.7× latency reduction (23.2s → 4.9s) with 0% accuracy degradation on engineering benchmarks

"9-agent stateful LangGraph platform for automated skill profiling, dynamic DAG learning path routing, and real-time mastery tutoring."

• Full production implementation: FastAPI backend, React 18 / Vite dual portals (Employee & Manager), Neo4j 5.15, Redis 7, and PostgreSQL pgvector
• 9 LangGraph agents: Competency Architect, Learning State Manager, Assessment Scoring, Content Generator (RAG), Content Reviewer, Learning Path Designer (Dijkstra/A* on a Neo4j skill graph), Adaptive Tutor (WebSocket), Mastery Evaluation, Orchestrator
• Multi-model inference: Dynamic routing across NVIDIA NIM (Meta Llama 3.1 70B & 8B Instruct), OpenAI, Anthropic, and Gemini
• 4-tier memory: PostgreSQL (metrics) + Neo4j (skill maps) + Redis (session state) + LangGraph checkpointers
• 🔗 Full Public Repository

"Secure, sandboxed filesystem bridge for AI coding assistants — built to connect internal AI assistants to codebases when commercial Copilot was restricted."

• 17 sandboxed developer tools: Multi-file batch reading, AST symbol extraction, function/class discovery, and ripgrep text search
• Robust security sandbox: Strict directory traversal prevention (.. blocking), symlink escaping checks (Path.resolve), and relative path validation
• Built on FastMCP v2 + Starlette / Uvicorn; auto-encoding detection via chardet; code metrics across 15+ programming languages
• 🔗 Full Public Repository

Actively seeking Automotive Systems Engineering × Applied AI roles internationally — Indian citizen, open to relocating worldwide, visa sponsorship required.

If you're building next-generation automotive R&D platforms — let's talk.

_{Built for safer, faster, and smarter automotive engineering.}