Extraction stuck at ~70%
Long regulated documents
Gold-set regression harness
Measure before tuning
Rebuilt evidence retrieval
Plus targeted model tuning
extraction accuracy
I build production LLM and agentic systems, and the secure multi-tenant backends under them, for regulated domains: legal, life sciences, fintech. Intelligence sits inside the workflow; every output is verified, cited and measured.
Hard-coded workflows break on the case nobody wrote. I put a model in the loop that reads each case, decides what to do, checks its own work, and hands off to a person when it isn't sure. Pick a document and run both.
Long, regulated documents. At every step the model makes a judgment, and every judgment is verified, cited to its page, and measured against a gold set.
What the agents do, and what changed.
Long regulated documents
Measure before tuning
Plus targeted model tuning
extraction accuracy
GCS → MongoDB
Two hard constraints
Batch processing
15% ahead of schedule · retrieval time −40%
Two clinical SaaS products
Queries and data model
Data modeling
API efficiency
Backend and data first, then retrieval and agents, and now the controls that let AI run in regulated work.
B.Tech, Chemical Engineering, Anna University (ACT Campus), Chennai · 2016–2020 · CGPA 8.19/10
4× Shining Star (monthly best performer), Saama · INSPIRE Science Camp, Govt. of India, top 200 nationally