// agentic ai · llm grounding · founder

Pradeep.

I build production AI systems that hold up when a customer is actually watching.

Engineering~0% → ~96% valid structured-output rate, from a single training/inference alignment fix

// about

From architecture
to production

I build AI systems end-to-end, from the architecture decisions to the code that ships. Currently building two products as an independent founder: an LLM-grounded optimizer for corporate treasuries, and Scorecery, an adaptive SAT-prep product live in soft-launch. Before that, I led engineering for an agentic AI platform at Oliv.ai, and spent 15 years at Cisco building cloud infrastructure and data platforms at scale.

What I care about is the gap between a demo and something that survives a real customer's stack, real data, real failure modes. Closing that gap is architecture, not luck.

Currently focused on

Agentic system architecture & SDKs
LLM grounding without hallucination
Fine-tuning & structured generation
GPU infra & inference optimization

// experience

Where I've worked

// projects

Selected work

Featured

Scorecery

A fine-tuned open model generates structured SAT items; a 5-layer automated eval gate (parse, schema, self-consistency, numeric-consistency, LLM-as-judge) scores and gates every item before a student ever sees it. Feeds a full digital-SAT-faithful adaptive exam app with IRT-based scaled scoring. Fixing a training/inference format mismatch was the unlock: it lifted valid structured output from ~0% to ~96% in a single alignment fix, the change that made every downstream stage possible.

Engineering ~0% → ~96% valid structured-output rate, from a single training/inference alignment fix

Engineering 5-layer automated eval gate scoring every generated item before it ships

Python FastAPI Fine-Tuning LLM-as-Judge IRT Scoring

View project →

retriever.py

# Eval gate: every generated item passes all five stages before shipping
gate = EvalGate([
  ParseCheck(),
  SchemaCheck(),
  AnswerSelfConsistency(),
  NumericRationaleConsistency(),
  LLMAsJudge(model="judge-model"),
])

result = gate.run(item)
if not result.passed:
    reject(item, reason=result.first_failure)

Scorecery — Adaptive Exam Engine

The full-stack exam app behind Scorecery: a digital-SAT-faithful 4-module adaptive test with IRT-based scoring, not just the generation pipeline that feeds it.

Python FastAPI IRT Scoring Adaptive Testing Integration Testing

View →

Intercompany Settlement Optimizer

Graph + LP optimization for corporate treasury cash repatriation, with an LLM-grounded analyst layer. Currently in stealth, invite-only demo coming soon.

Python FastAPI LangChain CrewAI LLM Grounding ILP Optimization

// writing