AI/GenAI engineer in Bengaluru. Currently Data Engineer II (MLE) at Redica Systems, working on document intelligence, retrieval agents and LLM features. Previously Software Engineer – ML at Zolvit (formerly Vakilsearch), working on legal-document search, summarisation and extraction.
I build independent testing and evidence for agent controls.
| Project | Question it answers | State on 12 October 2026 |
|---|---|---|
| actseal | Does a frozen action policy meet its declared risk and coverage limits? | 1.0.1 on PyPI; separate finite-benchmark audit: INCONCLUSIVE, collected with unreleased snapshot d3edbab; replay requires that archived producer |
| agent-reliability-ci | Did an agent change reduce reliability under tool/MCP faults, and what failed? | Evidence release 2026-10-06: planner, technical report, 18 archived decisions re-derived byte for byte on the tested macOS arm64 and Linux x86-64 environments |
| frontier-scout | Did an agent PR stay within the scope declared on its base branch? | 2.2.0 on PyPI; repairs the verifier defects published for 2.1.0 and earlier |
| dorian | Does the code still satisfy its recorded, executable claims? | 1.4.0 on PyPI as dorian-vwp |
| evalopt-graph | Do the supplied gates, claims and evidence satisfy an acceptance policy? | 0.1.0 on PyPI |
Try the synthetic demo. Requires macOS/Linux and uv; use a new ./actseal-demo
directory. The first command may download Python and the package.
uvx --python 3.12 --from "actseal==1.0.1" actseal demo --out ./actseal-demo # exits 0: BLOCK for the bad run, PASS for the fixed run
uvx --offline --python 3.12 --from "actseal==1.0.1" actseal replay ./actseal-demo/bad/evidence # exits 1 with the reasonOr start with ARCI's offline demo: in its seeded retry example, removing a retry changes success from 192/200 to 132/200; the gate reports BLOCK and the reduced failure replays offline (recorded results).
What the audits say, not what the READMEs promise. Actseal's 8 October audit of a live decision model is published with its verdict, its sealed inputs and the exact producer needed to replay it. On a fixed 16-intent Banking77 subset, the policy acted on 580 of 639 verification cases; 24 actions disagreed with benchmark labels. The risk CP interval [0.0250, 0.0639] crosses the 0.05 limit: INCONCLUSIVE. This is demo-scope evidence collected with unreleased benchmark snapshot d3edbab; replay requires that archived producer.
Frontier Scout's defect matrix lists the bypasses that 2.1.0 and earlier accepted, with the regression tests that now reject them.
How these were built. I use AI coding assistants as pair-programmers and record that assistance in commit trailers. I can walk through the problem framing, verification semantics, statistical decision rules, threat models and tests on a call.