Selected engineering work

The work is
open to inspection.

Public code and technical writing show how Nikhil approaches evaluation, failure handling and scope. Each example states what the evidence supports.

01 / Public repository / Evaluation

Evalharness

Read the repository

Deterministic checks for citation shape, injection patterns, personal-data patterns and lexical support. The repository includes synthetic fixtures, tests and mutation checks that make the method inspectable.

Evidence
Source code, test commands, synthetic examples and documented limitations.
Limit
The checks make no model calls. They are heuristics, not validation of an LLM judge or evidence of a client outcome.

02 / Technical case study / Agent memory

Shared memory with a provenance trail

Read the case study

A technical record of recall, handoffs and provenance across agents, including local verification and the boundary between a working local system and a production cutover.

Evidence
Published implementation notes and a local verification checkpoint.
Limit
Production identity and credentialed cutover were pending at that checkpoint. The case study does not establish production performance.

03 / Technical case study / Scope & control

A large agent design, reduced to a guarded change

Read the case study

A design and correction case showing how a broad multi-agent blueprint was narrowed to one guarded pull-request workflow.

Evidence
The published design decisions, corrections and implementation boundary.
Limit
A standalone runtime was not present at the documented checkpoint. This is a design and correction case, not a deployed platform claim.

Apply the method

Make your next change
easier to verify.

Bring the behavior you need to fix and the evidence you already have.

Book a 30-minute call