Skip to content

Manas Dasari

AI/ML engineering · research · applied AI internships, and freelance work

CSE undergraduate (5th semester, class of 2028), Amrita School of Engineering · Bangalore, Karnataka, India

I run machine learning experiments built to be checked (held-out tests, baselines, null results reported) and build the tools around them.

Selected evidence

  1. Clause Risk Review

    NLP research

    Does legal pretraining help a model find risky contract clauses?

    • BERT: 0.879 document F1 / 0.933 AP on 76 held-out contracts
    • Null result reported: Legal-BERT gave no measurable gain over BERT (one seed so far)
    • Leakage-safe: splits per contract, thresholds tuned on validation only, 90 tests

    read the case studycoderesults dashboard

  2. Med-EEG

    EEG research

    Can two EEG numbers track a meditator’s focus?

    • MNE pipeline: filtering, ICA, epochs, spectral and coupling features
    • Leave-one-subject-out, permutation tests and non-EEG baselines; the expertise result is confounded with sex, and says so
    • Pre-registered replication on a second dataset (39 people): not replicated; depth decoded at within-person AUC 0.609

    read the case studycodeNeuroDial dashboard

  3. Sprout

    Developer tool · v0.6.0

    Map your codebase, for you and your AI agent.

    • A single Go binary: gitignore-aware trees, git hotspots, PRs as trees, a suggested reading order
    • Token-budgeted project maps for LLMs, and an MCP server coding agents call directly
    • Released through a Homebrew tap, a Scoop bucket and Linux packages

    project pagecodelive site

More: every project, including Atmos, Forme and Flow.

Skills, and where they’re used

Engineering
Go · JavaScript / TypeScript · React · FastAPI · Git, Linux, CI · MCP

Open source, in other people’s projects

Writing