Applied AI, machine learning and process automation, built close to the operations they are meant to serve.
I have spent 25 years working through messy problems, bringing together people, process, technology and whatever else the situation needs. Twenty of those years were in research operations, which is where I learned how much of a system lives in its exceptions. I keep adding to that range through mathematics, software technology, AI and machine learning, and I stay close to the implementation because solving a problem often demands it.
This profile holds the code. The reasoning behind it, including case studies, working principles and longer arguments, lives at amitmohite.in.
- Applied AI and LLM systems. Context engineering, evaluation design, and deciding which judgments belong to a model, which belong to deterministic code, and which stay with a person.
- Machine learning in practice. Benchmarks that are able to fail an attractive idea, leakage-aware evaluation, and reporting what a result does and does not support.
- Process automation. Understanding what experienced people actually do before automating it, so the judgment that was protecting the work is not quietly removed with the clicks.
- Data and MLOps. Pipelines, experiment tracking, and the unglamorous work that decides whether a model survives contact with production.
Can a person's survey history be compressed into model weights well enough to predict their unseen answers? I built a leakage-aware benchmark on the public Twin-2K-500 panel and tested a low-cost identity-token recipe against a model-free floor and a prompted baseline. Recall of already-seen answers rose from 51.1% to 61.9% while held-out accuracy fell from 47.1% to 41.8%. The adapter was memorising the training table, not learning the person. The negative result is the useful part.
Python · PyTorch · LoRA fine-tuning · Qwen2.5-14B · Statistical evaluation
→ Read the full case study
A Python implementation of Random Iterative Method weighting, following Deming and Stephan (1940), that aligns survey sample distributions with known population margins. Convergence checks, weight bounds and diagnostics included, because it was written for real market research work rather than illustration.
Python · pandas · Survey methodology · Iterative proportional fitting
A translation can be linguistically plausible and still be wrong for a particular study, because what the text is for and who it speaks to change what it has to say. I treated context as explicit, reviewable data rather than hidden prompt text, gave exact rules to deterministic code and meaning to the model, and kept the decision with the person. Testing which context actually changed a judgment mattered more than collecting more of it.
Generative AI · Context engineering · Python · Structured evaluation · Human in the loop
→ Read the case study (source not public)
Most of my professional work sits inside confidential, long-running operations, so it cannot be published as source. I use that experience to explain how I think, and the write-ups at amitmohite.in/work go further into the reasoning than a repository can.
I write about applied AI, automation and the human side of technical change: what to understand before automating it, where context belongs in a system, and what responsibility looks like when a model is in the loop.
- Understand the work before automating it. Why useful AI systems start with how skilled people handle intent, exceptions and judgment, rather than with a model choice.
- Context is part of the system. Translation validation as a worked example of why more context is not automatically better.
- Designing transformation for agents, not only interfaces. An argument for exposing clean logic and data to intelligent agents rather than building one more screen.
→ All ideas · How I approach problems
| Area | Tools |
|---|---|
| Languages | Python, SQL, C#, .NET |
| ML and AI | PyTorch, scikit-learn, pandas, LoRA fine-tuning, evaluation harnesses, LLM and agent tooling |
| Platform | Azure ML, Microsoft Fabric, Airflow, MLflow, Docker |
| Practice | MLOps, data pipelines, model deployment, process automation, APIs |
A small reference set from formal study in mathematics, software technology, AI and machine learning. These are coursework rather than current work, kept public because the working is visible: melanoma detection (CNN) · hand gesture recognition (CNN) · banking ticket classification (NLP) · syntactic processing of medical text · telecom churn · bike sharing demand · lead scoring with Airflow and MLflow
- Site: amitmohite.in, for case studies, approach and ideas
- LinkedIn: linkedin.com/in/amitmohite
- Email: amit.mohite@outlook.com
Good problems deserve good conversations. If something here is useful to your work, I am happy to talk.