Open Source
Everything I build outside client work is public. This is the code behind the research papers, the agentic AI systems I have been building to understand the tooling properly, and the applied computer vision work — all of it readable, most of it runnable.
All repositories live at github.com/kiran-collab. Descriptions below are what each repository actually contains, not what it aspires to.
Agentic AI and LLM systems
Built to understand agent architectures end to end rather than through a framework's demo — including the parts that are usually skipped, like evaluation harnesses and CI.
Multi-agent MCP PDF Invoice Extractor
Invoice extraction built with Google's Agent Development Kit as an Orchestrator, File, and Extraction agent architecture, using Model Context Protocol for standardised file access. Pydantic validation and a trace-driven evaluation framework measure extraction accuracy, hallucination rate, and workflow reliability.
regsentinel
A multi-agent compliance auditing system built on the Claude Agent SDK, orchestrating five specialised agents against an MCP tool server to automate risk assessment and citation validation. Includes guardrails, LLM evaluation pipelines, and CI/CD with security scanning and test coverage.
Evaluation Pipelines in Agentic Frameworks
Evaluation harnesses for agent systems — the measurement layer rather than the agents themselves. Concerned with what "working" means for a multi-step agent, where the failure surfaces, and how to catch a regression that only appears three steps in.
AI Agents for Image and Video Generation
Agent workflows driving generative image and video pipelines — prompt construction, model selection, and the orchestration around a generation step rather than the model doing the generating.
Fitness & Nutrition AI Agent
A domain agent over nutrition and training data — a smaller, self-contained example of tool use and structured output where the domain constrains what a correct answer looks like.
fraud-det-plat
A fraud detection platform sketch — streaming feature serving, model scoring, and the surrounding plumbing, built to work through the architecture rather than to ship.
Code behind the research
Implementations from published and submitted work. Where the paper is the contribution, the repository is what makes it checkable — see Projects for the write-ups.
Griffin
Middleware framework for real-time road event detection using a UAV and a ground vehicle together — the coordination, data fusion, and alert path rather than the detector alone.
CSIKit-BMEye
BMI estimation from commodity WiFi channel state information, built on the CSIKit toolkit. Device-free sensing — the signal is the measurement, with no camera and nothing worn.
SWiDir
Walking direction estimation that corrects smartphone inertial sensing with passive WiFi, for indoor positioning where GPS is unavailable.
SAIND
Scene animation from RGB images and depth maps. Contributed to spsingh37/SAIND — collaborative work carried out in a fork.
Style Interpolation
Font style generation for low-resource Indic languages, improving scene text recognition by widening training diversity. Contributed to pantDevesh/Font-Interpolation.
Applied computer vision and ML
Self-contained implementations, mostly built to work a technique out properly rather than to read about it.
Deepfake Detection Browser Plugin
A browser extension for real-time deepfake monitoring and blocking — inference inside the browser, where the latency and packaging constraints shape the model choice more than accuracy does.
Face Recognition
The full pipeline — detect, align, embed, compare — rather than a wrapper around one library call.
Masked Image Inpainting
Reconstructing masked regions of an image — the generative task where the evaluation is harder than the model.
Sign Language Recognition
Gesture recognition over video — a sequence problem rather than a per-frame classification one, which is the point of building it.
Negative Comments Analyzer
Text classification over user comments — sentiment and toxicity, where the interesting problem is threshold choice rather than model choice.
TID Effect in Spiking Neural Networks
Total ionising dose effects on spiking neural networks — how radiation-induced hardware degradation propagates into model behaviour.
Numerical and exploratory
Work from outside the ML track, kept public because the implementations are useful to someone.
Inexact Inverse Iteration for Hermitian NLEPs
A numerical solver for Hermitian nonlinear eigenvalue problems, where the inner linear solve is deliberately inexact and the question is how much inexactness the outer iteration tolerates.
Quantum Bit Error Rate
QBER computation for quantum key distribution and efficient photonic transmission.
book-distribution
A Flutter application — included because it is public, and because not everything needs to be a model.
Upstream contributions
Related: Projects for the research write-ups, and Hackathons for time-boxed builds.