No. 01iOS game
Fold & Flip
Book cricket and book basketball for iPhone, built for the iPhone Duo hinge. Open the phone, a page appears, and its last digit is your ball.
- SwiftUI
- Core Haptics
- Game Center
KishoreLabs
a one-person workshop for games, apps & engines, all made by hand
I'm Nikhil Kishore, a senior machine learning engineer. For 8+ years I've built production AI, ML and distributed data systems: agentic LLM platforms, RAG and tool-calling workflows, on-device AI, and real-time infrastructure that moves 19 PB of streaming data and billions of events a day.
On my own time I ship AI products and open-source tools, mostly because I can't leave a good idea alone. The range runs wide, from memory systems that forget on purpose to habit apps and iPhone games you play with a hinge. Python, Rust, Swift, Dart, Java: whatever the idea needs.
Chapter 01
The Work
Five things off the bench. Each keeps its own notes, privacy policy and terms of use, so the App Store always has somewhere to point.
No. 02iOS app
Knot in Time
A habit tracker with no subscription, 100% offline. Goals, streaks, health logs and an on-device AI journal, now with widgets, Lock Screen timers and alarms.
- Flutter
- Foundation Models
- WidgetKit
No. 03memory engine
iforgot
ForgetfulDB, a forgetting engine and memory layer for AI agents, backed by SQLite. Decay forgets, salience keeps, habit reinforces, dreaming creates.
- Rust
- SQLite
- Ollama
No. 04Python library
Seedmill
YAML in, training data out. Config-driven synthetic data for LLM training sets, generated against local models or a hosted API.
- Python
- Ollama
- llama.cpp
No. 05iOS app
Offing
The social network for places where the internet isn't. A feed, chat and radar carried phone to phone over a Bluetooth mesh. No bars, no problem.
- Flutter
- Bluetooth LE
- Protobuf
Chapter 02
Experience
8+ years of applied AI, agentic systems and ML infrastructure
- 2023 — now
Agentic AI & ML platform
Comcast · Senior Machine Learning Engineer- Architected a multi-agent platform (RAG, context-aware discovery, tool calling) that turns a new telemetry endpoint from several sprints of work into minutes.
- Shipped Clarity: plain-language questions over huge telemetry and customer datasets, with multi-agent workflows on Google ADK and ClickHouse, and a Flutter client packaged as a native macOS app.
- Built TACOS, which routes context between agents using ant colony optimization to cut token use in production LLM pipelines.
- 2017 — 2023
Production ML & streaming
Comcast · promoted to Senior (Engineer 4) in 2022- Architected a real-time ML and telemetry platform on Apache Flink: 19 PB of streaming data and billions of events a day, at sub-second latency.
- Trained production deep-learning models for telemetry classification and anomaly detection.
- Built pipelines from Kafka through Flink and Spark into ClickHouse, Redis and DynamoDB, with retention and replay, and tuned ClickHouse to query billions of events a day.
- Owned the Java and Spring Boot ingestion services that keep every stream’s schema honest.
- nights & weekends
Kishore Labs
Independent studio · design, code, ship- Knot in Time and Offing are on the App Store; iforgot and Seedmill are open source. Seedmill generated the training data for Knot in Time’s on-device classifier.
The toolbox
AI & ML
- PyTorch
- TensorFlow
- scikit-learn
- LLM fine-tuning
- RAG
- Multi-agent systems
- Tool calling
- On-device inference
- Core ML
Data & ML systems
- Apache Flink
- Kafka
- Spark
- Databricks
- ClickHouse
- Redis
- DynamoDB
- MySQL
Languages & platforms
- Python
- Java
- TypeScript
- Swift
- Dart
- Rust
- AWS
- GCP
- Kubernetes
- Docker
- Spring Boot
- Flutter
- SwiftUI
Schooling
- MS, Applied Data Science & InformaticsNew York University2015 — 2016
- BSc (Hons), Computer Science, AI specializationBrunel University London2011 — 2014
Chapter 03
Writing
experiments, build notes and the odd strange idea, on Medium
- Jul 2026
What If Neural Networks Could Sleep? ↗
Sleep-inspired training experiments, and why switching dropout on only when overfitting starts gave the smallest train-test gap.
- May 2026
When the Loss Landscape Has a Map: Lessons from Hessian-Guided Optimization ↗
A curvature-guided escape that found the global minimum in two dimensions, failed at ten, and left two ideas worth keeping.
- May 2026
Knot in Time ↗
Building the app with an AI coding agent, what went wrong, and an on-device pipeline that turns journal text into trackers.
- Dec 2024
De-Aging Through Multidimensional Time ↗
A speculative essay that models aging as a mix of real and imaginary time, with a few films along the way.
- Sep 2020
Consume protobuf from kafka connector in Apache Flink ↗
A tutorial on reading protobuf messages from Kafka into a Flink job with a custom deserializer.
Chapter 04
The Fine Print
every app gets its own privacy policy & terms; these are the links Apple asks for
| No. | Project | Privacy policy | Terms of use | Updated |
|---|---|---|---|---|
| 01 | Fold & Flip | /fold-and-flip/privacy | /fold-and-flip/terms | [DATE] |
| 02 | Knot in Time | /knot-in-time/privacy | /knot-in-time/terms | August 19, 2026 |
| 03 | iforgot | Not written yet | Not written yet | — |
| 04 | Seedmill | Not written yet | Not written yet | — |
| 05 | Offing | offing.xyz/privacy.html ↗ | offing.xyz/terms.html ↗ | — |
Questions about your data? Each policy names who to write to, or reach Kishore Labs at kishore.nikhil2090@gmail.com.