Karthick Balaje Gopalakrishnan Elango kb2326

I build LLM agents that are evaluated before they ship and governed once they're live.

1.00
_{escalation recall
ClaimLens}

33 / 33
_{red-team attacks held
ClaimLens}

1.00 / 0.91
_{knockout recall / precision
Grant Capture}

0
_{invented facts in drafts
Grant Capture}

29 / 32
_{labelled failures fixed
Health Agent}

🚀 Now

In production: an LLM agent fleet for a US direct-to-consumer apparel brand. It runs 21 scheduled jobs across 10 commerce and marketing APIs, checks every claim against live data before it reaches the founder, and has a person approve every dollar of ad spend.
Just shipped: the Calibrated Health Agent is public: an agent on my own health data with a web app, an eval suite and a CI gate (below).
Research: a pre-registered study of Gemma 4 coding agents, public with its paper.

🫀 Calibrated Health Agent

An AI agent you talk to about your own wearable and journal data. It plans, writes SQL or Python, runs it in a sandbox, and every number links to a receipt you can re-run. A code check, not a prompt, flags any figure no tool computed.

One ReAct agent with thinking and 14 tools on Google ADK; specialists as tools
265 eval cases, my own labels on 110 answers, a red team and a CI gate on every PR
After error analysis and a redesign, 29 of 32 labelled failures answered better
Web app over AG-UI: streamed steps, tappable receipts, live-location weather, behind Google sign-in
Seven forecasters make sealed daily predictions, hash-chained and published
Runs on GCP (Cloud Run, Terraform) for under $1/month

Code · Write-up · Forecast scoreboard

Google ADK Gemini AG-UI Next.js Vertex AI Terraform

🧪 Asking vs. Adding a Stage

My paper for Kaggle's Gemma 4 Developer Agent competition (paper track): A Pre-Registered Test on a Gemma 4 Coding Agent. Gemma 4 31B fixes real GitHub issues within a fixed 4.5-minute budget; I compared prompt rules with a planning stage.

Designed on 30 tasks, then tested pre-registered claims once on 41 held-out tasks
Prompt rules are followed to the letter (20/20 vs 0/25), not in intent
A planning-stage gain seen in development did not replicate, and the paper says so
Every table recomputed from per-task results in a public notebook

Code · Reproduce the tables

Gemma 4 Google ADK vLLM Kaggle Pre-registration Paired statistics

🧩 Selected work

🚗 ClaimLens

Car-damage claims triage. My own vision models measure the damage, a LangGraph agent reasons over the evidence and the policy wording, and fixed rules decide. Only a person can deny a claim.

Damage segmentation mask mAP50 0.733 (course baseline: 0.137)
Escalation recall 1.00 · $0.010 per claim
33 of 33 red-team attacks held · eval gate in CI

Live showcase · Write-up

YOLO11 LangGraph MCP Hybrid RAG DVC

📑 Grant Capture Agent

Finds federal funding a small R&D firm can actually win, rules out the ones it can't (quoting the clause and page), and drafts proposals from the company's own evidence.

Knockout recall / precision 1.00 / 0.91
Drafts: 10 of 10 true gaps flagged, 0 invented
Every module built twice; ablations picked what ships, and misses are reported

Google ADK Gemini Corrective RAG pgvector MCP

📈 cordilla-account-scoring

Turns an existing scoring model into a rep's morning: who to call, what to refresh first, what to leave, and what to say.

Logs cost and data quality on every run
Watches for the slow drift that broke the last scoring rollout

Python Claude LangSmith

🏭 supplier-discovery-agent

Turns a free-text buyer request into a ranked, verified top-5 supplier shortlist.

Routes each part of a query to the right engine (vector, SQL, graph, external API)
Fuses results with RRF, scores them with MCDA, and verifies before ranking

Claude Vector search SQL Graph

🧰 agentic-tool-scaling

Design and prototype for an agent that works across 500+ tools: a two-stage semantic tool router, OpenAPI ingestion, and a hierarchical LangGraph orchestrator. LangGraph OpenAPI Semantic routing

Earlier work

feedback-analysis-dashboard: sentiment analysis, topic modeling and user segmentation (Streamlit, scikit-learn, Plotly)
Smart-Grid-ML-Optimization-Project: energy demand forecasting pipeline

🧭 How I build

Models measure, the agent reasons, rules decide. Anything with consequences goes through code a person can read.

Evaluate the evaluator. An LLM judge is only useful once you've measured how often it agrees with a human.

Build the outer loop first. Specs, evals, tracing and approval rules come before the clever part.

Publish the losses. Pre-register the test, report the baseline, and keep the result even when the simple rule wins.

🛠️ Stack


Agents
Evals & safety	Golden sets · LLM-as-judge with human agreement · CI eval gates · red-teaming · pre-registered experiments · proper scoring rules
Retrieval & data	Hybrid search (BM25 + vectors) · LanceDB · pgvector · corrective RAG · semantic layers for agents
Models	PyTorch · YOLO11 segmentation · TimesFM · calibration · scikit-learn
Shipping	FastAPI · Next.js · Docker · GitHub Actions · DVC · Terraform · OpenTelemetry · GCP (Cloud Run, IAP, Vertex AI, BigQuery)