I break down hard technical problems and design the simplest architecture that solves them. Recent work: serverless knowledge pipelines, edge task scheduling, and railway signaling security.
Engineering
/ResearchAt Alstom, I mapped trust boundaries in the R1 signaling system across a 4-zone security architecture.
Ran the Initial Risk Assessment under IEC 62443. Authored a 40-page technical specification covering 23 industrial risks across CBI, ATC, and ATS subsystems.
Selected engineering works.
Recall
A serverless pipeline that captures social media links and articles from mobile share sheets, extracts text and audio via Gemini and Whisper, and auto-syncs structured notes directly to Google Keep.
- -AWS Lambda OCI Container (Python 3.11 + stripped ffmpeg)
- -Amazon API Gateway HTTP API for low-latency webhook ingress
- -Amazon ECR Private Docker Registry for reproducible builds
10 carousel slides processed in a single Gemini Vision call with zero OCR text loss.
Sub-second speech-to-text fallback parsing video audio directly with static ffmpeg.
Pixel Pusher
A serverless image processing pipeline that uses pre-signed S3 URLs for direct client uploads and triggers asynchronous Lambda transformation with a Docker-compiled Pillow layer.
- -Amazon S3 Dual-Bucket Isolation (Raw Input vs Resized Output)
- -AWS Lambda triggered asynchronously via s3:ObjectCreated events
- -Amazon API Gateway HTTP API generating short-lived presigned URLs
Clients upload binary payload directly to S3 input bucket via temporary signed URL.
Decoupled output bucket ensures newly created thumbnails never re-trigger resizing Lambdas.
Research Paper RAG
A grounded academic document QA system parsing two-column scientific papers with PyMuPDF, indexing in-memory FAISS vector representations, and streaming answers via Groq Llama 3.3 with chunk attribution.
- -FastAPI Backend (Python 3.11) + React Vite Frontend on Vercel
- -HuggingFace Inference API (all-MiniLM-L6-v2 384d vectors)
- -Groq SDK running llama-3.3-70b-versatile for sub-second generation
- -In-Memory FAISS Vector Index (IndexFlatIP cosine search)
Block-level sorting prevents out-of-order text merging across academic columns.
Llama 3.3 returns precise paragraph citations, falling back to strict non-hallucination bounds.
Deadlines
A role-based appointment scheduling platform featuring granular weekly doctor availability definitions, calendar-specific exception overrides, and collision-free booking locks.
- -React.js + Vite frontend with React Router and Context API
- -Node.js + Express.js RESTful API layer
- -MongoDB Atlas database with Mongoose ODM modeling
Granular weekly schedule matrix combined with calendar-specific date overrides.
Discrete portals for Patients, Doctors, and Administrators with bcrypt password security.
I write papers on applied systems too.
Benchmarking lightweight YOLOv11n against YOLOv8n on INSAT-3D satellite imagery yields 0.921 mAP@.50-.95, while naive geometric data augmentation degrades detection accuracy.
Systematic evaluation across evolutionary, swarm, and physics-based metaheuristics identifies convergence bottlenecks in real-time IoT task offloading architectures.
Adapting human queuing search heuristics to multi-tier Cloud-Edge-Mist scheduling slashes task execution latency by 55% with a 99.9% task completion rate.