Agentic AIAgent Summoning
LangGraph supervisors, Google ADK, MCP tool use. Multi-agent pipelines with checkpoints, retries, and human-in-the-loop review.
Lakshya
Khandelwal
Lakshya Khandelwal
MS Data Science @ Columbia
· Graduating Dec 2026
✦ Scholar of the Arcane Data Arts · Columbia, Class of 2026 ✦
I build multi-agent LLM systems that do real work — and the evaluation, data pipelines, and guardrails that make them trustworthy in production. Summoner of agents, brewer of forecasts, and keeper of data pipelines. I bind language models into fellowships that answer questions, file tickets, and brew campaigns — with wards that keep them honest.
> ⁂ ▍
- 0+
- developers served by my LLM cost-analytics agentswizards served by my cost-scrying agents
- 28%
- MAPE, down from 28% — XGBoost over SARIMAforecast error, down from 28% — a sharper crystal ball
- 0M+
- records explored by my 4-agent analytics systemrecords scried by a council of four agents
- 0%
- RAG accuracy at 1/10th the cost after benchmarking 4 model familiesoracle accuracy at a tenth of the gold
ML & ForecastingDivination
Gradient boosting, time series, and statistical rigor. I care about the baseline you beat and how you know you beat it.
Data & PlatformsPotions & Pipelines
Scrapers to SAP, SQL to dashboards, FastAPI to Lambda and Cloud Run. The pieces that turn a notebook into something people use.
ExperienceThe Chronicles
Where I've shipped.Deeds recorded in ink that does not fade.
-
AI/ML Engineer Intern @ Calix
- Built a conversational infrastructure ticketing agent (Google ADK, Gemini 2.5 Pro, MCP) that resolves VM, storage, and network requests in Jira Service Management with a draft → review → confirm loop.
- Developed a multi-agent LLM cost-analytics platform for 1,200+ developers, joining LiteLLM spend with org hierarchy and routing natural-language questions to specialist agents to track AI adoption.
- Engineered backend APIs for an internal platform hosting 7+ agents: auth, agent routing, session management, and automatic Jira/Confluence token refresh.
-
Data Scientist @ Anheuser-Busch InBev GCC
- Replaced SARIMA with an XGBoost demand forecaster (lag + rolling features) at SKU × distributor level, taking MAPE from 28% → 11%.
- Built Python scraping pipelines feeding SAP with 50,000+ records/day, cutting manual collection by 70%.
- Led root-cause analysis with supply chain and ops teams: −18% stockouts, +12% on-time fulfillment.
-
Data Science Intern @ Anheuser-Busch InBev GCC
- Shipped three Power BI dashboards on automation KPIs and bot performance for 15 stakeholders.
- Automated ingestion and transformation pipelines (SQL, Python, cloud), removing 60% of manual hours.
-
M.S. Data Science @ Columbia University
Natural Language Understanding, Artificial Intelligence, and applied ML — with most of my evenings spent turning coursework into deployed systems (see below).
ProjectsEnchantments
All three are live. Click through and try them.Each one still breathes. Go on — cast them yourself.
Multi-agent · ContentA Fellowship of Twelve
Live ↗
BrandBuddy
One prompt in, a full Instagram + email campaign out. A 12-node LangGraph pipeline where a supervisor orchestrates trend discovery, generation, and publishing agents.
prompt→trends→generate→rubric eval↺publish
Rubric-based self-evaluation triggers automatic retries; checkpoints let a human step in without restarting the run.
Multi-agent · AnalyticsThe Council of Four
Live ↗
LENS
An AI analyst for 10M+ retail records that returns citation-backed insights, charts, and recommendations instead of just a SQL result.
SQL→EDA→hypothesis→supervisor↺ ×8
A 4-agent reasoning loop with up to 8 iterations, built-in data-quality checks, and statistical tests.
RAG · EvaluationThe Oracle of Ledgers
Live ↗
FinancialRAG
Ask questions about 15 public companies' filings and get cited answers. Token-based chunking, metadata filters, and ChromaDB vector search.
Benchmarked GPT, Claude, Gemini, and Llama to choose the model on cost and accuracy together. Deployed serverless with FastAPI on AWS Lambda.
Beyond the résuméLore & Legend
How I work, and what I'm after next.The unwritten parts of the tale.
/now✦ Presently…
- Final semester of the MS in Data Science at Columbia.
- Looking for full-time AI/ML engineering roles starting early 2027: agentic systems, LLM platforms, applied ML.
- Building more agent systems with honest evaluation built in.
Measure before you believeTrust no prophecy unproven
Every system I ship has a baseline it has to beat: SARIMA for the forecaster, four model families for the RAG bot.
Agents need guardrailsEvery summoning needs a ward
Draft → review → confirm. Rubric checks and retries. Checkpoints where a human can step in. Autonomy is earned.
Ship it so people use itA spell uncast is no spell
Every project here has a URL. I'd rather put a working v1 in front of someone than polish a notebook.
Talk to the people with the problemHeed the villagers
The 18% drop in stockouts came from sitting with supply chain and ops teams, not just from the model.
ToolkitThe Spellbook
What I reach for.Incantations, mastered and practiced.
LLMs & AgentsSummoning
LangGraphLangChainGoogle ADKMCPLiteLLMOpenAI SDKRAGText-to-SQLLLM EvaluationHuggingFace
ML & ModelingDivination
PyTorchTensorFlowKerasScikit-learnXGBoostTime seriesDeep LearningComputer VisionA/B Testing
LanguagesTongues
PythonTypeScriptSQLJavaRScala
Data & InfraPotions & Wards
FastAPIReactDockerKubernetesAWSGCPAzureSnowflakeSparkPostgreSQLMongoDBChromaDBPower BITableau
Let's build something.Send word by owl.
I'm open to full-time AI/ML roles and interesting collaborations. The inbox is the fastest way to reach me.Quests, collaborations, or a good riddle. My owl checks the inbox daily.
lk3068@columbia.edu