Ashman Das

I build tools that check what you can’t.

I’m Ashman Das, 17, in Greater Noida West, India. Voice agents, safety audits and ML research, all built on one laptop.

See the work Email me

EmovereJune 2025 VantageJuly 2026 GraceJuly 2026 to now CASCaDERAug to Sep 2026

Four things I built, and what each one showed me.

Each one is here with the part I’m proudest of and the part that isn’t finished.

  1. agentic_completion_rejected
  2. agentic_preexec_open_app
  3. agentic_read_pdf
  4. agentic_step_failed
  5. agentic_two_step
  6. agentic_unparseable_plan
  7. blind_app_vision_mode
  8. conversation_plain
  9. empty_transcript
  10. fastpath_adjust_volume
  11. fastpath_close_app
  12. fastpath_cua_list_windows
  13. fastpath_delete_file
  14. fastpath_lock_computer
  15. fastpath_open_app
  16. fastpath_open_calculator
  17. fastpath_open_file
  18. fastpath_search_files
  19. fastpath_tool_error
  20. followup_agentic
  21. followup_chain_three_turns
  22. rate_limited_intent
  23. rate_limited_planner
  24. repeated_action_escalates
  25. safety_confirm_accepted
  26. safety_confirm_declined
  27. safety_confirm_hijacked
  28. safety_press_key_confirm
  29. stale_field_replace
  30. unparseable_intent
All 30 recorded test runs. The 10 in red are named after a way Grace failed.

Grace

Voice control of Windows for people with motor disabilities. Say what you want in your own words. Nothing leaves the machine.

Building. Moving from Python to Rust.
  • NUREG-07001,818 passages
  • ISA-101202 passages
  • IEC 61508 and 61511rule profile only
  • EEMUA 191rule profile only
  • MIL-STD-1472Hrule profile only
How much of each standard Vantage can quote. Two have their text transcribed. The other three are checked against rules, without quoting.

Vantage

Checks control-room screens in industrial plants against five safety standards, and cites the clause behind every finding.

IIT Kanpur AETL capstone.
Answer loss falls from 0.0394 at 1 loop to 0.0167 at 4 loops, then stays flat at 8 and 16 loops. 0.020.030.04 124816 1 loop: loss 0.03942 loops: loss 0.02194 loops: loss 0.01678 loops: loss 0.016916 loops: loss 0.017 best
Loss on the easy task by number of loops, lower is better. After 4 loops nothing changes.

CASCaDER

Research on letting a language model decide how much compute each word deserves. Run on one laptop GPU.

Paused until February 2027.
Illustration: noisy per-turn emotion readings, and a smoother line that carries over between turns and fades. turn 1 turn 12 carried state one turn’s reading
An illustration, not data. Each dot is one turn’s reading. The line is what Emovere carries forward: it moves with the readings but fades instead of jumping.

Emovere

An offline voice that keeps someone company in a crisis until a person can reach them.

Proposal. The first prototype’s code is lost.

The same lesson, three times

Grace

AI agents drift off what you asked. The drift hurts most the people least able to correct it.

Vantage

A verdict is only worth the source it points to.

CASCaDER

One uncontrolled variable can cost you the whole experiment.

All three are the same failure: you find out too late, because finding out early was too expensive. Most of what I build now is an attempt to make finding out cheap.

Notes

All notes

Ideas

Things I haven’t built yet, written down with a date so they stop evaporating. A spin-gravity space station, a helmet that knows the way, and a script checker that explains instead of editing.

Read the ideas