Abhishek Kaushik

Applied AI engineer available for select work

I build agents
you can audit.

I design agentic systems that can reason, retrieve, act and explain themselves—from incident investigation to long-term memory and tool validation.

Currently a Cloud Developer at Hewlett Packard Enterprise, where I work across ReAct agents, LLM infrastructure and high-throughput systems. I build the rest in public.

Email me ↗abhibuilds.work ↗GitHub ↗LinkedIn ↗Resume ↗

Working thesis

The best AI products don't feel magical. They feel legible. You should know what happened, why it happened, and what to do next.

Selected work

2023—26
PROJECTOPEN SOURCE

Operational Intelligence Engine

Evidence-grounded incident investigation for production systems. An agentic investigation system that ingests Alertmanager alerts, resolves affected applications and gathers evidence across logs, deployments, GitHub and RAG sources.

Decisions

1Deterministic workflows own routing, tool execution, evidence collection and citation validation; the LLM only synthesizes grounded hypotheses.

2Structured investigation state, provider-agnostic tools and end-to-end traces make each conclusion auditable and reproducible.

3Supports Ollama, OpenRouter and Gemini with Redis caching and Prometheus observability.

Repository ↗
PROJECTOPEN SOURCE

Anveshan

Local-first evaluation and observability for AI agents. A trace evaluation system for multi-turn agents that keeps execution data local and makes failure modes legible.

Decisions

1Extends OTLP with session and conversation context for multi-turn agents.

2Scores prompt source—human, agent or system—so self-prompting agents are evaluated fairly.

3Makes hallucination taxonomy first-class: fabricated commands, invented APIs and unverified assertions are scored directly.

Repository ↗
PROJECTOPEN SOURCE

QueryMind-DW

A data warehouse you can question in plain English, with guardrails. Natural-language-to-SQL is easy to demo and risky to run. QueryMind measures generated queries and blocks destructive statements before execution.

Decisions

1Kafka ingestion, MinIO lake, dbt transformations, DuckDB and PostgreSQL serving, with Streamlit dashboards.

2Schema-aware prompting with metadata retrieval; measured at 67% accuracy across aggregations, joins, subqueries and window functions.

3Redis caching reduced repeated-query latency from 14.9 seconds to under 2 milliseconds.

4A SQL safety validator blocks destructive statements before execution.

Repository ↗
PROJECTOPEN SOURCE

Open-source packages

Two Python libraries published on PyPI. mcp-agent-kit scaffolds production-ready MCP servers with A2A agent cards and portable SKILL.md packaging. toolbroidge defines a tool once and runs it across OpenAI, Anthropic or Gemini with strict schemas and retries.

Decisions

1SageScan is a local-first data-quality CLI with 17 validator types, KS-test and PSI drift detection.

2SageScan benchmarks 125K+ rows/sec on chunked CSV with 599 MB peak memory on a 1M-row dataset.

mcp-agent-kit ↗toolbroidge ↗SageScan ↗

Applied AI in production

HPE · Sep 2025—present

Cloud Developer at Hewlett Packard Enterprise, Bengaluru.

Selected work

•Built long-term memory and skill-validator components for a ReAct agent architecture.

•Replaced legacy S3 batch ingestion with Kafka streaming, sustaining 50K records/sec across Protobuf, Avro and JSON.

•Designed an AWS Bedrock and LangChain audit pipeline for PII exposure and policy violations in infrastructure logs.

•Migrated batch ingestion to Delta Lake for ACID guarantees, upserts and schema evolution.

Open source

2021—25

Contributor to LlamaIndex, dbt-core and more, with multiple merged pull requests.

LlamaIndex #20661 · #20787 · dbt-core #12390 · #12358

Education

2019—23

B.E. in Artificial Intelligence and Data Science

Ramaiah Institute of Technology · Bengaluru