Himanshu DongrePune, India

Trying to understand intelligence by building it.

AI systems engineer working across production ML, intelligent systems, and applied research.

Portrait of Himanshu Dongre under flowering branches.

I build production AI systems and stay close to architecture, deployment, optimization, and the tradeoffs that only emerge under real constraints.

Outside work, I use small independent systems like Smriti, Sentinel OS, and PropOps to think more clearly about reasoning, agency, trust, and decision support.

Personal projectsBackground
12+ yearsproduction AI and software systems
14-20 engineersled across architecture and delivery
Portrait of Himanshu Dongre under flowering branches.

Current work

  • Reasoning workflows with explicit state and recoverable context.
  • Trust boundaries, policy, and verifiable execution for AI systems.
  • Decision-support tools for messy, incomplete, real-world evidence.
Experience

Experience.

The main line of my professional work.

2019 - PresentBengaluru / Tokyo / Pune

Sr. Tech Lead, AI/ML Autonomous Systems · KPIT Technologies

Led AI/ML systems work across perception, in-cabin sensing, and deployment while staying hands-on with architecture, optimization, middleware, and delivery.

  • Led teams of 14-20 engineers across perception and in-cabin AI programs, covering architecture, execution planning, technical reviews, and delivery.
  • Built multi-model, multi-sensor systems spanning camera, LiDAR, radar, pose, gaze, and anticipation components.
  • Optimized and deployed models on constrained platforms using ONNX, TensorRT, quantization, distillation, graph optimization, and inference acceleration.
  • Designed middleware for concurrent perception workloads and tuned a custom CenterNet pipeline to 33 FPS on full-resolution input.

Earlier work

  • Software Engineer · Independent / Freelance Work2017 - 2018 · Bengaluru
  • Software Engineer · Accenture2014 - 2017 · Bengaluru
Full background
Projects

Personal projects.

Current independent work. I keep this list intentionally narrow.

Reasoning infrastructure2026Active

Smriti

Checkpointed reasoning state for long-running, multi-model work.

reasoning stateagentsmulti-model workflows
Read notesRepository
Trust kernel for agents2026Active

Sentinel OS

A local-first trust kernel for agent systems with policy checks, append-only history, and verification.

trustworthy agentspolicy enforcementcryptographic verification
Read notesRepository
Decision support system2026Prototype

PropOps

A property-diligence system that gathers fragmented records, surfaces inconsistencies, and keeps judgment with the human operator.

decision supportretrievalrisk analysis
Read notesRepository
Interests

Questions I keep returning to.

Recurring questions behind both the systems I build at work and the ones I build independently.

Reasoning and state

How decisions, assumptions, and partial work are represented, restored, and compared across tools and models.

Trust and verification

How agent systems expose policy, evidence, and execution history in ways that can actually be checked.

Interfaces for intelligent systems

Where abstractions remain legible to humans instead of collapsing into opaque workflows.

Decision support under uncertainty

How systems help judgment without pretending certainty where the evidence is incomplete or noisy.

I am most interested in intelligent systems when they become concrete: how state is represented, what can be verified, where interfaces fail, and how useful behavior survives real constraints.

Approach

Working defaults.

A few defaults that shape how I build and evaluate systems.

  • First principlesThe work should still make sense after you strip away tooling, hype, and surface detail.
  • Implementation as a testBuilding is part of thinking. Weak assumptions usually show up faster in systems than in discussion.
  • Clarity over theatreI care more about what survives scrutiny, constraints, and deployment than what merely performs well in a demo.
Building to understand. Understanding to build.

I care about the loop between ideas and implementation. If an idea gets weaker once it touches code, constraints, interfaces, or data, it probably was not strong enough yet.

Log

Project log.

A short public record of releases and meaningful project changes.

Apr 2026Prototype

PropOps: early prototype.

A narrow public prototype for property diligence, centered on gathering fragmented legal and registry signals and surfacing inconsistencies for human review.

NotesRepository
Apr 2026Release

Smriti: public repo and demo.

Checkpointing, branching, and restoration for long-running reasoning state across sessions, tools, and model switches.

NotesRepository
Feb 2026Milestone

Sentinel OS: first public milestone.

A clearer public direction for local-first policy enforcement, append-only execution history, and verifiable agent behavior.

NotesRepository
Links

Contact.

The best entry points are the projects, repositories, and direct email.

Reach out or follow along

Email is the easiest way to reach me. You can also find current public work through GitHub, LinkedIn, X, and Instagram.

EmailDownload CV (PDF)

Public links

GitHub is the best place to start for current public work. LinkedIn and X carry occasional public updates and context.