Three internships, three disciplines. Descriptions are sanitized, and reconstructions are invented rather than copied from employer material.
Pass 1 of 3OverviewThree roles, one line each 1 minute
No internship is tagged with this discipline yet. See Valerie and PHAROS for agent work.
2026
Ultraviolette Automotive
ML Engineer Intern · Bengaluru · June 25 to August 19, 2026
I built speech processing that runs on the device and analysis that links every finding to its evidence, with numerical checks left to deterministic code rather than the model.
I worked on the teleoperation backend and the operator interface for driving a robot remotely, including the hand-offs between autonomous and manual control.
◆ Owner confirmed All three roles, their dates and their figures are from my November 2026 résumé. They are my own account: nothing here was re-measured for this site, and each figure links to its line in the evidence ledger.
Pass 2 of 3How it worksWhat each role built, with the reconstructions and figures 3 minutes
Ultraviolette Automotive: what it built
ML Engineer Intern · Bengaluru · June 25 to August 19, 2026
Reconstruction: extraction versus verdictOpen
▦ Reconstruction with invented data
Fictional scooter, invented specification and synthetic dialogue. A reconstruction of the idea, not company data. A model extracts a claim from language; deterministic code decides the verdict.
Claims about the fictional Scooter S1
Said
Extracted
Specification
Verdict
“It does about 140 on a charge.”
Range ≈ 140 km
Range: 120 km
Contradicted Deterministic comparison
“Top speed is 90.”
Top speed = 90 km/h
Top speed: 90 km/h
Matches Deterministic comparison
“It charges really fast.”
Nothing extracted
Charge time: 4 h
Not checked Recall failure: the verifier cannot repair a missed extraction
Built edge transcription/diarization and evidence-linked analysis with deterministic specification verification.
Sanitized public summary. The built engine is distinct from the designed backend and app.
Benchmarked on-device faster-whisper (INT8) transcription and pyannote diarization, reaching 96.5% speaker attribution on 50 hand-annotated recordings of real showroom speech.
Owner-described scope; no engineering source material was reviewed.
Pass 3 of 3Deep diveWhere each figure comes from, and the rooms not built yet as long as you like
Every figure above links to its line in the evidence ledger, with its status and caveat. The reconstructions are invented, never copied from an employer.