N.V.J.K Kartik
AI engineer at Future AGI, in Bangalore. I pick a field and go all the way down: classroom speech, conversation analytics, medical imaging, markets, and now evaluating AI agents.
First author, AgentCompass (arXiv) and a T4E 2024 chapter · Co-author, IEEE TALE 2024 Best Student Paper · Co-inventor, US 12,608,610
Every work, as a list ↓Drag to look around. Click the floor to walk there, or a work to go to it.
WASD walks; the arrow keys do too once you’ve clicked in.Tap a work or the floor to walk there. Drag to look.
Plan of the room
From above: the entrance at the bottom, the newest work nearest the door.
Start here
A harness that watches the productError FeedGrouping failed AI runs into issues and investigating each one with an agent, fast enough for tables with millions of spans.
A shared AI gatewayAgent Command CenterKeeping provider-specific tools intact when every provider goes through one door.
A migration under a live featureAnnotations on ClickHouseTwenty PRs moving annotations and Error Feed onto ClickHouse before their Postgres tables were dropped, without failing open or running out of memory.
Everything I’ve worked on, newest first
Jun–Jul 2026
- Annotations on ClickHouseI was responsible for the annotation side of the migration.
- Error FeedI built the investigation agent, kept the feed alive through the ClickHouse migration, and made it fast at scale.
Posts about what brokePosts on evals, tracing and agents in production, most of them starting from something that broke.
Dec 2025 – Apr 2026
Synthetic data generation
AI Evaluation 1.0
Simulate SDKSep–Oct 2025
Agent Optimizer
Triggers of shared regulation
AgentCompassJun 2024 – Jun 2025
The open-source SDKsReleases I shipped or co-authored.
traceAII wrote traceAI’s TypeScript SDK in May 2025, and its C# and Java SDKs for v1.0.0, which I published in March 2026.
Speaker diarization in classroomsCo-author, second of seven. My part was the diarization pipeline work.
Market sentimentI ran transformer-based sentiment analysis over social media and financial news.
Dec 2023 – Apr 2024
Nexus
Centio.AI
Alumni ConnectAug–Dec 2023
Energy-efficient lung nodule analysisDesigned models that were about 77% more energy-efficient, with an 86% cut in compute use and no loss in accuracy.
AffectBotsI built AffectBots as part of the research group’s classroom AI work.
Conversation analyticsI built speaker diarization and voice emotion recognition models and the React app people used to run them.
Feb–Apr 2023
Emotion detector for online classes
HivemindTechnical writing
Posts about what broke.
- My CI evals were green. A regression still paged me at 3 AM.Medium · 3 min · 17 Sept 2026
- LlamaIndex makes RAG easy to build and hard to debug. Here is how I evaluate it.DEV · 6 min · 2 Sept 2026
- Why Your Agent Passes Every Eval and Still Fails in ProductionMedium · 8 min · 2 Sept 2026
- My DSPy pipeline compiled beautifully and got worse in productionDEV · 3 min · 31 Aug 2026
Rabbit holes, not hobbies.
An agent, a classroom, a lung scan, a market, a brain: I open the black box and poke at the mechanism until it makes sense. At work, so far:
I’m Kartik. I build things across research and product, and I like the stretch where they meet: a model that has to survive real users, or an app that needs real ML underneath it.
At Future AGI I work across the pieces a team needs once its agent is live: tracing (traceAI), evaluation, simulation, Agent Optimizer, the Agent Command Center gateway, and the ClickHouse data and annotation systems underneath. They are one loop: see what the agent did, score it, rehearse it, improve it, route it, and find out what broke. Most of my time goes into that last step: AgentCompass, then Error Feed.
In December 2023 I was at three of the places below at once, plus the degree.
The thread through all of it: build the thing, then check that it actually works. The posts are about the times it didn’t.
The path
- Dec 2024 — nowFuture AGIAI engineer: intern from Dec 2024, full-time from Jul 2025
- Jun 2024 · 3 monthsIIT Madras · RBCDSAIResearch Intern
- Dec 2023 — Jan 2024NIT PuducherryMLOps Intern, Dept. of CSE
- Aug 2023 — May 2024Vocab.AIFull Stack Developer Intern
- May 2023 — Aug 2024IIT BombayResearch Associate, Educational Technology
- 2021 — 2025IIIT DharwadData Science & AI
- Built the Error Feed investigation agent, which reads a cluster of failing traces and returns a cause, a fix, a confidence and its evidence, and kept the feed alive and fast through the ClickHouse migration.
- Rewrote the evaluation SDK for 1.0: one evaluate() call routed between local metrics, Turing models and LLM judges.
- Wrote traceAI’s TypeScript, Java and C# SDKs, and published v1.0.0 in four languages.
- Started Simulate SDK for voice agents and added chat; published the Agent Command Center SDK 1.0.
- Moved the annotation system’s reads onto ClickHouse before the Postgres tracer tables were dropped.
- First author of AgentCompass; co-inventor on a granted US patent.
- Transformer-based sentiment analysis over social media and financial news.
- Combined sentiment with technical indicators and fundamentals, and studied how feature selection and time-series models change prediction quality.
- Designed energy-efficient models for lung nodule classification and segmentation on LIDC-IDRI, using pruning, quantization and architecture changes.
- Built speaker diarization and voice emotion recognition models, and the React app people used to run them.
- Fine-tuned Llama 2 and GPT-3.5 Turbo for sentiment, topic modelling and conversation scoring.
- Wired an automated QA system into customer service workflows, plus the REST APIs behind the analytics.
- Built AffectBots, a multimodal tutoring system that reads audio, video and text to recognise emotion in real time.
- Implemented speaker diarization for code-switched classroom audio with pyannote and Whisper. The paper won Best Student Paper at IEEE TALE 2024.
- Applied Conditional Random Fields to detect triggers of socially shared regulation in group problem-solving. First author on the T4E paper.
- Built Hivemind, a study-together web platform, with friends from class.
- Worked on Alumni Connect, a Flutter app for the college’s career centre.
- Hackathons: an emotion detector for online classes (DRS) and Nexus at Hackfest ’24.
Off the clock, currently
- NeuroscienceCurrently reading Behave; the “okay, but why does this happen?” side of biology and psychology.
- MarketsValuations, market mechanics, and the history of financial scams.
- StartupsBusiness models, unit economics, and “could this actually be a company?”
- HardwareGPUs, chips and compute economics. Adult LEGO, except one brick costs ₹3 lakh.
- F1Ferrari and Leclerc, and the strategy calls more than the zoom. Cricket and football too.
- PhilosophyFree will, consciousness, and why people do what they do.
Got a rabbit hole for me?
n.v.j.kkartikGitHub ↗LinkedIn ↗DEV ↗Medium ↗Hugging Face ↗npm ↗PyPI ↗
Based in Bangalore. Receipts for every fact on this site.