Satvik Adyanthaya — I build systems that make hard calls defensible.Ibuildsystemsthatmakehardcallsdefensible.

Problem

Every spring a student organization loses its leadership and most of its memory with them. The budget history survives as a file nobody can find; the reasoning behind it does not survive at all. The next person starts from zero, and the institution above them has no view across any of it.

Approach

• 01

  An AWS-backed platform with a relational model designed so the record and the reasoning behind it live together and outlast whoever entered them.

• 02

  A KPI framework and event workflows with role-based reporting, so an organization and the institution above it act on the same governed data.

• 03

  User interviews and mapped decision flows drove the schema; adoption, approval, and continuity signals are instrumented rather than assumed.

Outcome

First place and $2,500 in non-dilutive funding at Startup Wednesday Rochester, hosted by NextCorps, ahead of four other finalists. Now running an institutional pilot spanning roughly thirty organizations.

Problem

Text-to-CAD is where language models stop being impressive. “A bracket with four M6 holes on a 40mm circle” has to become geometry that is valid, manufacturable, and parametrically editable — not a plausible-looking mesh. Most systems produce output that fails the moment a kernel tries to build it.

Approach

• 01

  The model plans rather than draws: it emits a structured geometry programme, so every result is something a kernel can actually execute.

• 02

  Execution is deterministic. Confining the probabilistic step to planning is what makes a validity rate reportable instead of anecdotal.

• 03

  Verification runs on the produced solid, not on the text — a result only counts when the geometry itself checks out.

Outcome

Valid geometry on all 81 benchmark cases, and a No. 1 global ranking at the time of submission; it currently sits at No. 2. Also a ClawComp finalist at $5,000.

Problem

Biomedical hypothesis generation is still someone reading 200 PDFs and writing a memo. The useful cross-disciplinary connection sits buried across sources that share no vocabulary — and a system that surfaces one but cites it wrong is worse than useless to a researcher.

Approach

• 01

  Eleven specialised agents over a hybrid memory: a Neo4j graph for entities and ontology edges alongside a ChromaDB vector store for passages, so retrieval can follow a relation or a resemblance.

• 02

  PubMed-BERT handles domain embeddings rather than a general-purpose model — most of the difference between a citation that holds and one that merely looks right.

• 03

  Citation accuracy is the headline metric, measured against researcher review rather than asserted.

Outcome

Won the Concept Award at the NYBPC FLX Region, and holds 94% citation accuracy under testing with three researchers. The accuracy number is the point: in this domain, a system that cannot be trusted to attribute is one nobody uses twice.