Sena Köse@kosesena
4
Events attended
3
Submissions made
Turkey
1 year of experience
About me
Software engineering student and ABAP intern, building at the intersection of enterprise systems and AI. I'm interested in the unglamorous part: making an agent safe enough to write to a system of record. Currently building GlovesOn — a voice agent that posts goods receipts into SAP without the worker taking their gloves off.
🤓 Latest Submissions
The Felt Five: a factory that tests its own tests
The checks that ship with each stage are a sample of the judges' suite: 79% of stage 1, 9% of stage 3. Green on them proves little. The Felt Five is a factory that measures its own evidence: a spec-auditor breaks one requirement at a time in a throwaway copy and counts how many breaks the band's evidence catches, against how many the shipped checks catch. In the submitted pocketful run, across four stages, it seeded 81 faults. The band caught 75; the shipped checks caught 8, all of them among those 75; the other 6 change nothing a test can observe. The implementer's own tests caught 71 of the 75, so the result does not rest on the auditor's probes. A sixth seat, the cross-auditor, runs on Codex and audits every revision blind to the other verifiers. In the submitted run it alone rejected five revisions that all three Claude verifiers had accepted, among them a fractional amount rounded into a valid payment and a paid request that a hand-edited import allowed to be paid twice. It is not one-way: the customer seat rejected two screens the reviewer had accepted. The mandates are generic. With the same six files and only the dispatch changed, the factory also built the other scored track, tablekeeper: all four stages in 2 h 39 min, one human message, every stage folder claiming its stage in an isolated run. Every design choice is an ADR with its options measured, and FACTORY.md lists every action the operator took outside the room, including a restart script replaced during the run and an Xcode licence accepted on the host. Stage 4 closed with one recorded limitation, an import of about 300 older saved statements that exceeds the 5 s limit. Repo: github.com/kosesena/dark-factory-pocketful. Live demo: felt-five-pocketful.onrender.com (demo account [email protected] / felt-demo-ada).
6 Oct 2026
GlovesOn — voice goods receipts into SAP
The problem. A receiving clerk has gloves on and an unplanned pallet in front of her. Recording it means stopping, walking to a terminal and typing. GlovesOn lets her do it by talking. The workflow. She names the material and the quantity. GlovesOn looks up the description, unit and bin itself, reads the receipt back (quantity, unit, material description, bin) and posts it only after an explicit "yes". A wrong receipt is corrected with a reversal document; both documents stay. The same receipt twice is refused by a duplicate guard in the gateway. Built on AssemblyAI. The Voice Agent API carries speech, conversation and voice. Tool calls return to the browser and pass through a scoped capability to our gateway, which speaks SAP's released OData APIs with the CSRF handshake a real S/4HANA requires. The write tool does not exist until a draft has been read back. Honest limits. Reads have been verified against SAP's S/4HANA Cloud sandbox; writes go to a mock S/4HANA that follows SAP's contract, and nothing has been posted to a real SAP system. The posting is made by a service user, not the individual worker. Calls, emails and notes are demo records. Noise robustness is not benchmarked. Every claim, the file it lives in and the test that holds it: docs/JUDGE-GUIDE.md in the repository.
30 Sep 2026
Clean Core Copilot: IBM Bob reviews old SAP code
Problem. SAP ends mainstream maintenance for ECC 6.0 on 31 December 2027. Picture the one ABAP developer at a mid-size manufacturer with 600 custom programs and that deadline. Every company moving to S/4HANA Cloud carries hundreds of custom ABAP programs written over fifteen years. Each one must be checked against SAP's Clean Core rules and often rewritten before it can move. Today that review is manual, done by the few consultants who know both the old and the new world. It is slow, expensive and inconsistent, and nobody knows how far an AI assistant can be trusted with it. That is the question a project manager actually asks: not "can Bob do this?" but "how well, measured, and where does it fail?" Solution. Clean Core Copilot is an IBM Bob custom mode plus an evaluation harness. Before the hackathon we wrote four synthetic legacy programs (finance, sales, materials) with 41 planted problems, a twelve-rule Clean Core catalogue, and an answer key. The key stays outside the repository; its SHA-256 was committed on 24 September, before Bob saw any code, and the organisers confirmed the preparation. During the hackathon Bob, in the "Clean Core Architect" mode, audited all four programs and wrote JSON findings with a rule ID, line range and reason for each. We scored them against the sealed key and confirmed every judgement by hand. Results. Bob found 27 of 29 catalogue violations on the right lines and 1 of 12 business-logic defects. Of the 32 findings we could judge, 29 were correct (91 %): two were wrong and one partly right; six more were duplicates. Bob then rewrote one program into a CDS view, a class and seven ABAP Unit tests, and rewrote the other three in parallel with three sub-agents in sixteen minutes. On a real SAP BTP trial system the view failed on exactly the two SAP objects Bob had marked candidate; with two labelled stand-in tables the class compiled, Bob's seven tests passed and ATC reported zero errors. A differential test against the old rule on 431 cases passed too (8 of 8), and changing one limit from 30 to 31 days made two tests fail. Getting there exposed two syntax defects in Bob's code that only a compiler finds. What it means. Bob is a strong first reviewer for catalogue rules and a weak one for business logic. The human stays in the loop, and we can now say where, with numbers. A recorded-evidence site shows every finding next to the code, whether Bob was right, and what was and was not verified. Total cost: 4.26 of 40 Bobcoin. Limits. The samples are synthetic (customer ABAP cannot be published); the numbers describe this test, not Bob in general. One of four rewrites was tested, against stubs; three are unreviewed. Nothing was verified on a customer's system.
27 Sep 2026
