Building Reliable Intelligent Systems

1. Method ¶

Keywords below are used in the spirit of RFC 2119. The requirements are self-imposed, and I fall short of them more often than I would like; they are written down so that falling short is something I can see. They remain the most stable thing on this website.

• Understanding MUST precede implementation. I do not build what I cannot derive. Reading the paper is part of writing the code.
• Claims MUST ship with their uncertainty. A number without an error bar is an opinion.
• Systems SHOULD assume their own failure. Fault tolerance is designed in, not retrofitted on.
• Simplicity is REQUIRED; cleverness is OPTIONAL. The best architecture is the one a stranger can operate.
• Optimization without measurement is NOT RECOMMENDED. Profile first. The bottleneck is rarely where intuition points.
• Work not written down SHOULD be treated as lost. Hence this site: a numbered series of working notes, revised like code.
• The domain MAY change. The standard MUST NOT. Astrophysics yesterday; trustworthy AI, ML systems, or distributed computation tomorrow. The method travels.

2. Evidence: Engineering ¶

Research code that runs only on its author's machine is a rumor.

The current chapter is industrial. At iGPT Agency I build multi-agent LLM systems: an AI filmmaking platform that orchestrates twelve-plus agents across four generative APIs, replacing a manual prompt-and-regenerate workflow with a single gated pipeline and cutting per-scene production time by 70–85%. Alongside it, a candidate-intelligence platform that turns scraped profiles into ranked, embedding-scored shortlists for under a cent per role.

Before that, infrastructure: during the SparQ programme at QNu Labs I designed Kubernetes-based orchestration for scalability and fault tolerance, with deployment automated through Docker and CI/CD. Infrastructure at a security company is not allowed to fail quietly; building under that constraint taught me more about reliability than any course.

This section stays at abstract altitude by design. As the record grows, each significant build becomes a numbered note of its own (§6); what accumulates here is the index, not the detail.

where

iGPT Agency, Bengaluru

role

AI Engineer

period

Jan 2026 – present

surface

12+ LLM agents · 4 generative-AI APIs · embeddings · Playwright

result

70–85% per-scene production time cut · 90%+ screening time cut

write-ups

planned, see §6

where

QNu Labs, Bengaluru (quantum-security infrastructure)

role

SparQ Summer Intern

period

May 2025 – Jul 2025

surface

Kubernetes · Docker · CI/CD · distributed resilience

3. Evidence: Research ¶

The first large-scale validation of the method came from a signal-reconstruction problem. An instrument observes a decaying signal and keeps looking away, through orbital occultation, slewing and downtime, precisely when the signal matters most. Interpolating across the gaps quietly corrupts every parameter fitted downstream.

At the National Astronomical Observatory of Japan, I designed attention-gated U-Net architectures that learn the structure of the decay instead of smoothing over it, and investigated time-aware Neural ODE formulations that treat irregular sampling as a property of the data rather than a defect [1,2]. Validated across 521 instances, the approach reduced reconstruction-induced parameter uncertainty by 38–41%.

The signals were gamma-ray-burst light curves. The problem, inference under missing data with honest uncertainty, is not astronomical at all.

where

National Astronomical Observatory of Japan, Tokyo

role

Student Research Intern

period

Jan 2025 – Dec 2025

validation

521 instances, full catalogue

result

38–41% reduction in reconstruction-induced uncertainty

output

2 international publications [1,2]

write-ups

WN-002, Reconstruction Under Missing Data · WN-003, A Model That Did Not Work

4. Toolchain ¶

Tools, grouped by layer. The list changes slowly; the layers do not.

5. Direction ¶

Some problems keep pulling me back: trustworthy AI, reliable machine-learning systems, efficient inference, distributed computation, formal reasoning about software.

The common question, how do we build intelligent systems whose behaviour we can actually trust and how do we know, belongs to none of these fields and outlasts all of them.

I admire people who move between research papers and production systems without changing clothes. That is the career under construction here.

6. Planned Notes ¶

The notes themselves live in the indexes below rather than on this page, so this section stays the same length however many of them accumulate. Each index lists its own notes, newest last, and each note carries its own number and address.

§ References ¶

1. [1] Manchanda, A., Kaushal, A., Dainotti, M. G., Gupta, K., Deepu, A., Naqi, S., Felix, J., Indoriya, N., et al. “Gamma-Ray Burst Light-curve Reconstruction: A Comparative Machine and Deep Learning Analysis.” The Astrophysical Journal Supplement Series, 281:35, December 2025. Open access. doi:10.3847/1538-4365/ade851
2. [2] Kaushal, A., Manchanda, A., Dainotti, M. G., Gupta, K., Nogala, Z., Madhan, A., Naqi, S., Kumar, R., Oad, V., Indoriya, N., et al. “Multi-Model Framework for Reconstructing Gamma-Ray Burst Light Curves.” Journal of High Energy Astrophysics, 51:100519, March 2026. doi:10.1016/j.jheap.2025.100519

7. Author's Address ¶

email

nikunjindoriya@gmail.com

orcid

0009-0005-5663-4728

scholar

Google Scholar

code

github.com/nikunj-indoriya

profile

linkedin.com/in/nikunj-indoriya

elsewhere

x.com/indoriya_nikunj · leetcode.com/u/nikunj_indoriya

timezone

Asia/Kolkata (UTC+05:30)

Nikunj Indoriya received the B.S. degree in computer science and engineering, with a minor in data science and engineering, from the Indian Institute of Science Education and Research, Bhopal, in 2026. He worked on reconstruction under missing data at the National Astronomical Observatory of Japan, contributing to two international publications, and is an AI engineer at iGPT Agency, where he builds multi-agent language-model systems. His interests are trustworthy machine learning, uncertainty quantification, and the infrastructure that has to carry both.