Rickesh T N

CV and robotics engineer. 15 years in production software, 9 taking deep learning from research to deployed systems. I write about evaluation and training data, and the ways both quietly fail.

Singapore Joined on Aug 6, 2026 rickesh.t.n@gmail.com https://linkedin.com/in/rickeshnatarajan

Education

M.Sc. Digital Media Technology, Nanyang Technological University

Pronouns

He/Him

Work

Senior Computer Vision Engineer, Applied Research at KARL STORZ

Rickesh T N

Oct 16

I built an oracle to prove my benchmark needed full document coverage. I only tested it from one end.

#benchmarks #evaluation #llm #testing

5 min read

Want to connect with Rickesh T N?

Create an account to connect with Rickesh T N. You can also sign in below to proceed if you already have an account.

Create Account

Already have an account? Sign in

Rickesh T N

Sep 5

I told the model to separate fields with <TAB>. It did exactly that, and I lost 79 percent of my data.

#llm #prompting #debugging #localllm

3 min read

Rickesh T N

Sep 5

Temperature 0 is not reproducible. I measured 30 percent of my output changing between identical runs.

#llm #benchmarks #localllm #reproducibility

3 min read

Rickesh T N

Sep 5

Your agent truncates the corpus and answers anyway. Two harnesses, and a router that picks between them.

#llm #agents #localllm #benchmarks

5 min read

Rickesh T N

Sep 5

A small model on a 6GB laptop matched direct reading on aggregation, except when the answer is a number

#llm #benchmarks #localllm #agents

4 min read

Rickesh T N

Sep 5

Our small model beat Claude Opus on one 440K-token question. Then it came last on the public benchmark.

#llm #benchmarks #localllm #privateai

5 min read

Rickesh T N

Aug 24

My inference server decided my second GPU no longer exists. Here is how I got it back without upgrading a driver.

#gpu #llamacpp #ollama #homelab

3 min read

Rickesh T N

Aug 24

A generic fine-tuning playbook, written after doing it wrong several times

#llm #finetuning #machinelearning #mlops

6 min read

Rickesh T N

Aug 24

A small model on a 6GB laptop beat Claude Opus on one 440K-token question. The fix was giving the model less to do.

#privateai #llm #inference #localllm

5 min read

Rickesh T N

Aug 23

99% token accuracy, zero learning. Field notes from fine-tuning vision models with RL.

#machinelearning #ai #llm #debugging

4 min read

Rickesh T N

Aug 11

Your multi-agent system isn't hitting prompt cache. Your system prompt is the reason.

#llm #ai #architecture #performance

6 min read

Rickesh T N

Aug 9

Can a Cheap Model Beat a Frontier Model? Rebuilding Recursive Language Models with Codex

#ai #llm #agents #machinelearning

6 min read

Rickesh T N

Aug 7

We made our world model smaller and it got better. Then "efficient" attention made nothing faster.

#machinelearning #computervision #ai #performance

5 min read

Rickesh T N

Aug 7

Three Ways Your Training Data Lies to You (And None of Them Throw an Error)

#machinelearning #datascience #mlops #ai

5 min read

Rickesh T N

Aug 7

Your reasoning model isn't dumb. Your parser is throwing away its best answers.

#machinelearning #llm #evaluation #ai

4 min read

DEV Community

Rickesh T N

Badges

1 Week Community Wellness Streak

Writing Debut

Skills/Languages

Currently learning

Currently hacking on

Available for

I built an oracle to prove my benchmark needed full document coverage. I only tested it from one end.

Want to connect with Rickesh T N?

I told the model to separate fields with <TAB>. It did exactly that, and I lost 79 percent of my data.

Temperature 0 is not reproducible. I measured 30 percent of my output changing between identical runs.

Your agent truncates the corpus and answers anyway. Two harnesses, and a router that picks between them.

A small model on a 6GB laptop matched direct reading on aggregation, except when the answer is a number

Our small model beat Claude Opus on one 440K-token question. Then it came last on the public benchmark.

My inference server decided my second GPU no longer exists. Here is how I got it back without upgrading a driver.

A generic fine-tuning playbook, written after doing it wrong several times

A small model on a 6GB laptop beat Claude Opus on one 440K-token question. The fix was giving the model less to do.

99% token accuracy, zero learning. Field notes from fine-tuning vision models with RL.

Your multi-agent system isn't hitting prompt cache. Your system prompt is the reason.

Can a Cheap Model Beat a Frontier Model? Rebuilding Recursive Language Models with Codex

We made our world model smaller and it got better. Then "efficient" attention made nothing faster.

Three Ways Your Training Data Lies to You (And None of Them Throw an Error)

Your reasoning model isn't dumb. Your parser is throwing away its best answers.