ONFINANCE AI / INTERNAL WORK01
Document intelligence
Vision → document tree → agent traversal
§1 Overview
§2 Financial statements
§2.1 Revenue table
PAGES → SECTIONS → BOXES
HIMANSHU KUMAR / BENGALURU, INDIA
AIEngineer.
I build AI agents, retrieval systems,
and the infrastructure that runs them.
AI Engineer at OnFinance AI
IIT Kharagpur · M.Sc. Economics
Experience & impactIN PRODUCTION AT ONFINANCE AI Context & responsibilities ↗
100M+LLM tokens processed
per month
per month
20+Enterprise client
deployments
deployments
$260K+Infrastructure credits
secured
secured
01 / ENGINEERING IN PRACTICEFrom unstructured data
From unstructured data
to trustworthy agents.
Document understanding, agents that act on real systems, and the harness that makes them measurable. Explore the architecture and my contribution to each project.
Browse all six projects02 / PROJECT INDEX
Explore GitHub ↗Selected projects.
§1 Overview
§2 Financial statements
§2.1 Revenue table
PAGES → SECTIONS → BOXES
DOCUMENT INTELLIGENCE 01
Hierarchical Document Parsing
Built at OnFinance AI: a vision-based parsing engine that maps documents into pages, sections, and bounding boxes for targeted agent traversal.
INTERNAL Built at OnFinance AI. No public repository.
AGENTgathers facts · proposes
POLICY + APIpure functionapproverUNIQUEauditdecides · re-checks
80% ACCURACY · 5% FALSE REFUNDS
PRODUCTION AGENTS / EVALS 02
Refund Desk
An AI agent that resolves refund requests against a real backend. Policy, permissions, and idempotency stay in deterministic code, and an eval harness measures false refunds, not just accuracy.
~ python cli.py "count the files"
0tool_use · list_files1.2k
1tool_use · read_file ×21.9k
2end_turn2.7k
↳ context growth 2.1×
AGENT ENGINEERING 03
loop-lab
A hand-written agent loop for the Claude Messages API with no orchestration framework: harness guards, per-turn traces, verified prompt caching, and tests that run without a network.
polymarket · server.py MCP
How could this event move my portfolio?
MODEL CONTEXT PROTOCOL 04
Polymarket MCP Server
An MCP server that gives agents Polymarket tools: semantic market search, live order books, price history, and ARIMA probability forecasts, used from a LangGraph ReAct client.
AGENT WORKFLOWdurable · streaming
SANDBOX VMfsshellgitdev serversnapshot · resume
THE AGENT IS NOT THE SANDBOX.
OPEN SOURCE / CLOUD AGENTS 05
Open Agents
My extension of Vercel’s Open Agents reference app adds a MongoDB connector with read-only queries, execution limits, and redacted audit logs.
FORK Built on vercel-labs/open-agents (MIT). My fork adds a hardened MongoDB connector to the Python agent.
STREAM / ANOMALY DETECTIONSIMULATED
Signal Anomaly
ILLUSTRATIVE SIGNAL
STREAMING MACHINE LEARNING 06
Real-time Anomaly Detection
A streaming ML prototype using River and SNARIMAX to detect anomalies in simulated time-series data, with a Dash dashboard and CSV export.
Showing all 6 projects.
03 / EXPERIENCE
AI engineering, in production.
JUN 2025 — PRESENTAI Engineer CURRENT
OnFinance AI · Bengaluru
AI Engineer CURRENT
OnFinance AI · BengaluruBuilding AI agents for risk, audit, and compliance in financial services.
- Agent systems. Architected Agent Studio with LangGraph, a connected file system, a dynamic tool registry, and skill-based prompt routing.
- Retrieval at scale. Built LLM and document pipelines processing 100M+ tokens per month, including clause-level retrieval over SEBI, RBI, and IRDAI regulatory data.
- Production ownership. Owned AWS Kubernetes infrastructure and delivery across 20+ enterprise accounts, spanning SaaS and on-premise deployments. Operated MongoDB, Redis, Qdrant, RabbitMQ, and Trino with GitHub Actions and Argo CD.
- Business impact. Secured $260K+ in AWS and Cloudflare infrastructure credits; supported a $4.2M Pre-Series A through technical demos and product presentations.
MAY — JUN 2025LLM Trainer
Outlier · Remote
LLM Trainer
Outlier · RemoteRLHF preference ranking and pairwise comparison of model responses across coding and reasoning tasks, with evidence-based justifications for every rating.
- Authored structured feedback and failure-mode annotations used to shape reward-model signals for downstream reinforcement learning.
- Reviewed code generation for correctness, efficiency, and adherence to task specifications.
OCT — DEC 2024LLM Trainer
Soul AI
LLM Trainer
Soul AIEvaluated and benchmarked multi-model LLM outputs across STEM and non-STEM domains for factual accuracy, reasoning, and response quality.
- Audited agent tool-calling behaviour, validating function-call schemas, parameter correctness, and API integration reliability in code-execution traces.
- Engineered prompts with zero-shot, few-shot, and chain-of-thought techniques to improve reasoning consistency.
MAY — JUL 2024AI Intern
Schneider Electric · Bengaluru
AI Intern
Schneider Electric · BengaluruBuilt a GenAI knowledge bot for developers, using retrieval-augmented generation over internal documentation and codebases to return source-grounded answers and code references.
- Prototyped GenAI across knowledge management, automated testing, and code assistance, evaluating LLM integration patterns and embedding-based retrieval.
- Designed semantic indexing and vector search over internal documentation to improve discoverability of engineering knowledge org-wide.
SEP — DEC 2023AI Research Intern
Indian Institute of Technology, Kharagpur
AI Research Intern
Indian Institute of Technology, KharagpurResearch internship at IIT Kharagpur, alongside my M.Sc. in Economics.
HIMANSHU KUMAR / BENGALURU04 / BACKGROUND
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, from raw data and model behaviour to agent orchestration and the infrastructure underneath. The goal stays the same: build something useful that works in the real world.
A little outside the terminal.
Photography and filmmaking. Former Head of Photography and Advisor at IIT Kharagpur’s Technology Filmmaking and Photography Society.
IIT Kharagpur
M.Sc. Economics · 2020–2025
7.9 / 10.0 CGPA
AI4ICPS Certificate Programme
M.Sc. Economics · 2020–2025
7.9 / 10.0 CGPA
AI4ICPS Certificate Programme
05 / TOOLS I WORK WITH
Tools for the whole journey.
01
Intelligence
LLMs, agents & retrieval
02
Machine learning
Models, training & evaluation
03
Infrastructure
From deployment to operations
04
Data systems
Context, storage & communication
05
Applications
Interfaces & analysis
Explore an AI workflow
FROM REQUEST TO RESPONSEINTERACTIVE DEMO
02 / ORCHESTRATION
LangGraph connects reasoning, tools, and context into configurable agent workflows.
Select a node to explore
An illustrative workflow, inspired by my engineering work.
06 / WRITING
All posts on LinkedIn ↗Notes on the AI landscape.
I write on LinkedIn about where AI is heading and what it takes to build with it. Below are my five most-read posts on the subject, by LinkedIn impressions in September 2026, and more to explore.
The most important AI in India isn’t a chatbot. It’s a translator.
Bhashini and NITI Aayog are putting voice-first AI into government services across 22 scheduled languages: speech recognition, IndicTrans2 translation, and speech synthesis. Frontier models get the headlines; this is the layer that decides whether any of it reaches a billion people.
Read on LinkedInApple’s “Illusion of Thinking” and the limits of reasoning models
Given the exact solution algorithm, reasoning models still collapsed on hard puzzles. One executed 100+ Tower of Hanoi steps but failed four moves of an unfamiliar river-crossing puzzle: a reality check on memorisation versus generalisation.
Read on LinkedIngpt-oss: OpenAI’s first open-weight models since GPT-2
120B and 20B models under Apache 2.0, with tool use and MXFP4 quantisation; the smaller one runs in 16 GB of RAM. A welcome pivot, with caveats: text only, and no training-data disclosure.
Read on LinkedInDeepSeek’s harness makes even the agent loop a plugin
Codex and Claude Code build file tools into the harness core. DeepSeek, inspired by Cordis, treats every capability as a plugin, so tools can be swapped at run time without restarting the agent.
Read on LinkedInGoogle I/O 2025 and the cost of a universal assistant
Gemini 2.5, AI Mode in Search, Project Astra, Veo 3, and Android XR glasses, and the questions they raise about autonomy, privacy, and over-reliance on AI.
Read on LinkedInExplore more writingShow fewer posts 08
Jev: a decision engine for what LLMs shouldn’t do
Agents often force LLMs to act as routers and guardrails. Jev scores a finite set of typed choices in parallel instead of generating text, so there is nothing to parse and the output space is bounded.
Read on LinkedInThe OpenAI Navier-Stokes controversy
Two mathematicians spent a year on a problem; OpenAI says a model resolved aspects of it in 88 hours. What the dispute over research direction, private tool data, and credit says about AI inside research.
Read on LinkedInInside LinkedIn’s “360brew” ranking model
A centralised AI ranking system that gates reach in stages, weighs comment depth and dwell time, and suppresses templated, generic posts.
Read on LinkedInVeo and the future of video creation
Google’s Veo generates 1080p video beyond a minute long from a text prompt: faster prototyping and B-roll for creators, and a higher bar for original work.
Read on LinkedInSarvam AI’s full-stack platform for Indian languages
Sarvam 2B, trained on 4 trillion tokens across 10 Indic languages, the Shuka audio model, and voice agents priced at ₹1 a minute.
Read on LinkedInStrawberry, Orion, and the push toward reasoning
OpenAI’s reported reasoning model, built with Self-Taught Reasoner techniques for maths and programming, and what a successor could mean for reliability.
Read on LinkedInThe $1 Chevy: prompt injection in customer-facing bots
Dealership chatbots were talked into “agreeing” to sell a $58,000 Tahoe for $1. Why customer-facing LLMs need input validation, grounding, and guardrails.
Read on LinkedInLangGraph Studio, an IDE for agents
Visual graphs, live state editing, and debugging for multi-step agent logic, from the LangChain team.
Read on LinkedIn07 / RECOGNITION
Beyond engineering.
Institute Order of Merit SOCIAL & CULTURAL
Recognised by IIT Kharagpur’s Technology Students’ Gymkhana for distinguished contribution to social and cultural life on campus.
HAVE AN INTERESTING PROBLEM?
Let’s build
something useful.
I’m open to conversations about AI engineering, intelligent products, and the systems behind them.