Namana Manohar  ·  Bengaluru, India
HelloEnglish

I build AIproducts thatshow their work.

I work across product, systems and AI, from deciding what to build to measuring whether it actually works. Most recently I took Lethe from an empty repository to a live product in seven weeks.

What I'm looking for

Roles: solutions engineering, forward-deployed engineering or technical product.

What I'd do there: sit with your clients, turn what they need into a working demo or integration, and stay until it's actually used.

Where: London, Singapore, Dubai, Berlin, Dublin or the US. I'd need visa sponsorship, and can start as soon as it's approved.

How I work with people

Listen first

At INNATE I held 50+ conversations with retail contacts before building, then cut 3D for a 2D prototype so testers could use it sooner.

Watch the real workflow

For Dispatch Planner I watched how my family's delivery day actually runs, and built the app around that instead of a feature list.

Measure after launch

On Lethe, usage data showed where new users were leaving during onboarding. I rebuilt that step and measured again.

The journey

Studying2019 – 2023

B.E., Computer Science & EngineeringGlobal Academy of Technology (VTU), Bengaluru

Where the road starts: algorithms, databases, networks, and my first machine learning paper.

BuildingAug – Nov 2022

Full-stack web internshipsCompsoft (Aug – Sep 2022) · Varcons (Oct – Nov 2022)

First real code that other people used: front ends, APIs and databases.

Research2023

Paper in IJRARCo-author

A survey of datasets and machine learning and deep learning methods for assessing osteoporosis risk from spine MRI.

Founding2023 – Feb 2026

INNATEFounder · AI virtual try-on for clothing

  • Held 50+ conversations with retail contacts, possible co-founders and early hires.
  • Built a 2D computer-vision prototype and cut 3D, so testers could use it sooner.
  • Pitched accelerators and investors, and turned down an early seed offer to finish my M.Tech.
Studying2024 – 2026

M.Tech, IoT & Sensor SystemsVellore Institute of Technology

  • Thesis: camera-based robot navigation using ROS2, SLAM and YOLO.
  • Co-authored a paper and poster (Wiley proceedings, 2025) on a ZnO-coated optical fibre sensor for detecting methanol adulteration in ethanol.
ROS2SLAMYOLOsensors
Award2025

Winner, Best Chemical HackYantra Central Hack, VIT · Team lead

  • Led the team on a chemistry problem outside my field, and learned the chemistry of self-healing coatings within the 40-hour event.
  • Built a transparent self-healing coating using microcapsules.
  • Won Best Chemical Hack; the jury recommended the coating for a patent.
BuildingFeb – May 2026

Design MeHelp people make expensive decisions they won't regret

A decision engine, starting with a bedroom on a budget. The model reasons about needs; plain code owns every price and size. Each pick shows its evidence, how sure it is about each claim, and the alternatives it rejected.

Building · Founder, SoulBotsJun 2026 – now

LetheYour AI conversations, turned into a profile any AI tool can read

ExampleI'm moving the launch to the first week of November, payments aren't ready. I always write specs before I code.
ExampleDraft the email to Priya? She likes short updates, numbers first. Should I learn Rust? Not sure.

Tap a claim to see the words it came from.

  • Built end to end with AI coding agents: specs and decision records first, then code, review and tests.
  • An MCP server lets Claude and Cursor read a profile one scope at a time. Every read is logged.
  • Benchmarked against Mem0 and Zep on the same model with one blind judge, and published every result.
  • A private Vault: encrypted on your own device, with a language model running in the browser on WebGPU.
  • Found the onboarding step where new users were leaving, rebuilt it, and measured again.
98.1% grounded, 268 claims7 weeks to live2,500+ tests

Grounded means every claim Lethe made was traced back to the sentence it came from, measured on my own data with a separate judge model. The tests are automated checks that run on every change.

Next.jsTypeScriptPostgresMCPLLM evaluationWebGPU
BuildingSep 2026

Dispatch PlannerRoute planning for my family's delivery business

I watched how their day actually runs first. Now they enter today's drops and the app says how many people to send on which routes, with map links and lists ready for WhatsApp.

ReactSupabaseroute clusteringmaps
Building · SoulBotsSep 2026 – now

The LeagueAI agents take on real work; a judge model scores it

A glowing golden stag, one of the League's creatures
A jade horned sage holding a scroll, one of the League's creatures
  • Designed 66 creatures across nine elements, each with its own work strengths and weaknesses.
  • Every rule is measured before it ships. When one move won too often, I changed the rule.
  • The judge had to tell good work from bad on known examples before it scored anyone.
Next stop2026

Your team?

Looking for solutions, forward-deployed or technical product roles in London, Singapore, Dubai, Berlin, Dublin or the US. I'd need visa sponsorship, and can start as soon as it's approved.

namana.manoharmr@gmail.com

Technical

Code
Python, SQL, TypeScript
AI
LLM apps, RAG, agents and MCP, evaluation, PyTorch
Stack
Next.js, Postgres, Supabase, Vercel, Git
Robotics
ROS2, SLAM, YOLO

Working with people

Before
Discovery calls, demos, specs
During
Short build and measure loops
After
Usage data, onboarding fixes, written updates

Languages

Telugu
Native
English
Fluent
Kannada
Fluent
Hindi
Fluent
Mandarin
Conversational
Sanskrit
Tritiya exam

Hobbies