AI ENGINEER / BENGALURU, INDIA
Intelligence,
engineered for
production.
I’m Himanshu. I build AI agents, intelligent applications, and the infrastructure that takes them from idea to production.
01 / IMPACT AT ONFINANCE AI
Numbers that made it
all the way to production.
Figures come from my profile. They describe systems I built, shipped, and still operate.
02 / THE PIPELINE — STAGE 01
A request enters
the graph.
A request arrives carrying three things: the task to perform, the context around it, and the documents it is allowed to reach. Everything downstream is shaped by what lands here.
- Task
- Context
- Documents
02 / THE PIPELINE — STAGE 02
Orchestration decides
what happens next.
LangGraph wires reasoning, tools, and context into configurable agent workflows. At OnFinance AI I architected this layer as Agent Studio: a connected file system, a dynamic tool registry, and skill-based prompt routing.
- LangGraph
- Skill routing
- Tool registry
- Connected files
02 / THE PIPELINE — STAGE 03
The branch: retrieval,
tools, inference.
Retrieval
Vector search pulls the document context that actually matters into the workflow.
Qdrant · FAISS · rerankingTool calls
A dynamic registry hands the agent exactly the capabilities its current task needs.
Dynamic registry · connected FSInference
Model access for reasoning and generation, routed through a single gateway.
LiteLLM · AWS Bedrock · Groq02 / THE PIPELINE — STAGE 04
Out the other side:
something usable.
The graph returns a structured result an application or a person can act on — validated, typed, and accompanied by the reasoning trace that produced it.
- JSON
- Pydantic validation
- Reasoning trace
An illustrative workflow drawn from my engineering work. It is not live telemetry.
03 / WORKING STACK — LAYER 01
Intelligence
LLMs, agents & retrieval
- Python
- LangGraph
- LangChain
- LlamaIndex
- LiteLLM
- AWS Bedrock
- Groq
- Prompt engineering
03 / WORKING STACK — LAYER 02
Machine learning
Models, training & evaluation
- PyTorch
- TensorFlow
- Keras
- Scikit-learn
- XGBoost
- LightGBM
- Transformers
- spaCy
03 / WORKING STACK — LAYER 03
Infrastructure
From deployment to operations
- AWS
- Kubernetes
- Docker
- Linux
- GitHub Actions
- Argo CD
- Cloudflare
- Bifrost
03 / WORKING STACK — LAYER 04
Data systems
Context, storage & communication
- SQL
- MongoDB
- Qdrant
- Pinecone
- ChromaDB
- Redis
- RabbitMQ
- Trino
- MySQL
03 / WORKING STACK — LAYER 05
Applications
Interfaces & analysis
- FastAPI
- Pydantic
- TypeScript
- Swift
- Streamlit
- Dash
- Pandas
- Power BI
04 / SELECTED WORK
Built. Shipped. Open.
Six things I put into the world. The next six planes take them one at a time — or jump straight to one.
DEVELOPER TOOLS / SWIFT 01
TerminalHistory
Your terminal sessions, remembered. A native macOS menu bar app and CLI that captures sessions automatically and replays them with a fresh shell in the original working directory.
- Captures sessions through a login-shell wrapper across terminals, IDEs, Docker, and SSH.
- Day-grouped menu bar dropdown plus a dedicated search window.
- Stores sessions locally in SQLite, with no third-party package dependencies.
- Swift
- macOS
- SQLite
- CLI
AI & DATA / CONVERSATIONAL RETRIEVAL 02
CSVChat
A conversation with your CSV. Tabular data becomes searchable knowledge, so people can ask questions in plain English and get grounded answers.
- Loads CSV data, chunks it, and generates sentence-transformer embeddings.
- Stores embeddings in FAISS for similarity search.
- Uses a LangChain conversational retrieval chain with a Groq model for contextual Q&A.
- Python
- LangChain
- FAISS
- Groq
SYSTEMS & ENGINEERING 03
System Design Academy
A TypeScript and Next.js learning platform that takes developers from system-design foundations through distributed systems, operations, and interview preparation.
- Eight phases, from foundations and low-level design to distributed systems and reliability.
- Interactive simulations, multiple quiz formats, progress tracking, and an AI learning assistant.
- Case studies covering messaging, ride matching, video streaming, and search autocomplete.
- TypeScript
- Next.js
- System design
STREAMING MACHINE LEARNING 04
Real-time Anomaly Detection
Make the unusual visible. A dashboard that simulates a live data stream, detects unusual observations as they arrive, and lets you inspect and export the results.
- River’s Predictive Anomaly Detection with a SNARIMAX time-series model.
- Simulates concept drift and occasional anomalies, visualised with Dash and Plotly.
- Start, stop, and reset controls, plus CSV export of detected anomalies.
- Python
- River
- Dash
- Plotly
OPEN SOURCE / CLOUD AGENTS 05
Open Agents
Coding agents that live in the cloud. A durable workflow drives an isolated sandbox VM through tools, so the agent and the machine it works on stay separate.
- Runs chat-driven coding agents with file, search, shell, and task tools over a cloned repository.
- The isolated VM snapshots and resumes while the durable workflow streams and stays cancellable.
- My fork adds a hardened MongoDB connector: schema discovery, read-only queries with size and execution guards, rejected write operators, and a PII-redacting audit log.
FORK Built on vercel-labs/open-agents (MIT).
- TypeScript
- Next.js
- Sandbox VM
- Python
- MongoDB
DOCUMENT INTELLIGENCE 06
Hierarchical Document Parsing
A document as a tree, not a blob. Vision calls infer section boundaries and describe every node, so an agent can walk the structure instead of re-reading the page.
- Parses documents into a hierarchy of pages, sections, and bounding boxes, using vision calls to infer boundaries across varied layouts.
- Describes every node in natural language, so sections are found by meaning rather than coordinates.
- Adds tree-navigation tools — parent, child, sibling, label search, box filtering — and caches tree state, cutting vision API calls by more than 60%.
INTERNAL Built at OnFinance AI. No public repository.
- Python
- Claude Haiku 4.5
- Vision
- Agentic traversal
05 / EXPERIENCE — CURRENT
AI Engineer
JUN 2025 — PRESENTOnFinance AI · Bengaluru
Building AI agents for risk, audit, and compliance in financial services.
- Architected Agent Studio: a LangGraph workflow engine with a connected file system, dynamic tool registry, and skill-based prompt routing, orchestrating AI pipelines across multiple product verticals.
- Engineered LLM, RAG, and document intelligence pipelines processing 100M+ tokens per month, using LiteLLM, AWS Bedrock, OCR, FusionQuery search, and re-ranking.
- Built clause-level RAG pipelines and agentic workflows over SEBI, RBI, and IRDAI regulatory data and compliance circulars.
- Owned AWS Kubernetes infrastructure and client delivery across 20+ enterprise accounts, including NSE, BSE, Axis Finance, BharatPe, and PayU Finance.
- Secured $260K+ in infrastructure credits through AWS and the Cloudflare Startup Program.
- Contributed to a $4.2M Pre-Series A from Peak XV Partners through technical demos and executive product presentations.
- LangGraph
- AWS Bedrock
- Kubernetes
- RAG
- Argo CD
05 / EXPERIENCE — BEFORE THAT
How I got here.
- MAY — JUN 2025
LLM Trainer · Outlier, Remote
RLHF preference ranking and pairwise comparison of model responses across coding and reasoning tasks, with evidence-based justifications for every rating. Authored failure-mode annotations that shaped reward-model signals.
- OCT — DEC 2024
LLM Trainer · Soul AI
Evaluated and benchmarked multi-model LLM outputs for factual accuracy and reasoning. Audited agent tool-calling behaviour — function-call schemas, parameter correctness, API reliability — and engineered zero-shot, few-shot, and chain-of-thought prompts.
- MAY — JUL 2024
AI Intern · Schneider Electric, Bengaluru
Built a GenAI knowledge bot using retrieval-augmented generation over internal documentation and codebases to return source-grounded answers, and designed semantic indexing and vector search to make engineering knowledge discoverable org-wide.
- SEP — DEC 2023
AI Research Intern · IIT Kharagpur
Research internship alongside my M.Sc. in Economics.
06 / THE ENGINEER BEHIND THE CODE
An economist’s lens.
An engineer’s instinct.
I studied Economics at IIT Kharagpur. Today I work at the intersection of business and AI: understanding the problem, designing the system, and getting it into production.
I enjoy the whole journey — raw data, model behaviour, agent orchestration, and the infrastructure underneath. The goal stays the same: build something useful that works in the real world.
Photography and filmmaking. Former Head of Photography and Advisor at IIT Kharagpur’s Technology Filmmaking and Photography Society.
Awarded by IIT Kharagpur’s Technology Students’ Gymkhana for distinguished contribution to social and cultural life on campus — the Institute-level tier, above Honourable Mention and Special Mention.
M.Sc. Economics · 2020–2025 · 7.9 / 10.0 CGPA
AI4ICPS Certificate Programme
07 / WHAT’S NEXT?
Have an interesting
problem? Let’s build.
I’m open to conversations about AI engineering, intelligent products, and the systems behind them.
© 2026 Himanshu Kumar · Bengaluru, IN