AI Engineer / Agentic Systems / Retrieval & Memory / Full-Stack Products
Somay Kousis / Agentic Systems
I build stateful multi-agent systems, retrieval and memory pipelines, and bounded authority for AI agents. The work I care about is the part that holds up under contention, cost limits, and being wrong, not the demo that runs once.
Architected the orchestrator routing 10,000+ daily tasks across specialized subagents at RYSE, cutting misrouted payloads 33% and dispatch latency 24%.
Kept 25 of 25 concurrent writes on a contended counter, with 58 SERIALIZABLE conflicts caught and auto-retried, where a flat-file analog kept 1 and silently lost 24. Ships as a runnable benchmark, not a screenshot.
Cut CI log payload 76% to 93% before any model call, measured against Podman's own logformatter fixtures and reproducible from a fresh clone with no token and no API budget.
A company I am building now, contract work on Standout, orchestration at scale at RYSE, and client work delivered end to end. Each one is listed with the part that was actually hard.
Co-founder & CEO
AI writes more and more of the code that ships, and review is where it quietly breaks: someone approves a diff they never read.
aye aye makes the person shipping a risky change explain it out loud, and a separate judge grades that explanation against the real diff.
Boring changes are skipped on purpose, so the check never turns into the next rubber stamp.
Pre-launch with the waitlist open, building it with Charlotte Liu.
Software Engineer, Freelance
Shipped Standout Network, a live, paying LinkedIn networking product, from an empty repo in 7 days as the sole engineer: web app, Chrome extension, Python agent service, and Postgres backend.
Orchestrated 8 LangGraph agents with heavy/lite model fallback and a grounding check that drops any draft line it cannot trace to a source.
Engineered the extension around LinkedIn's modal editors and lazy-loaded tabs, fixing half-loaded reads that scored a 54 profile as 22.
Red-teamed production with 5 adversarial reviews and live probes, landing 21 fixes (row-level security, edge rate limits, prompt-injection guards, PII redaction in AI traces), backed by 136 agent tests and 38 DB suites.
AI Engineer Intern
Built the orchestration layer that routes work across a large registry of specialized subagents.
Engineered 5+ production-grade subagents using UV and LangGraph, implementing persistent memory layers for long-running task continuity.
Architected a central orchestrator agent to route 10,000+ daily tasks across specialized subagents, cutting misrouted payloads 33% and dispatch latency 24%.
Software Developer, Client Project
Delivered a multi-role recruitment product end to end for a local client.
Architected a Next.js and Supabase recruitment platform supporting 3 user roles with distinct workflows and role-based access control (RBAC).
Automated hiring logistics via custom interview scheduling pipelines, reducing administrative overhead by 40% (saving ~15 hours/week).
Deployed via Vercel with integrated real-time analytics, ensuring 99.9% availability for a seamless onboarding experience.
B.Tech, Computer Science and Engineering
Aug 2024 – April 2028 / Gwalior, IN
Winner (International). Led a 4-person cross-functional team as technical lead, driving product direction.
2nd position (Hardware Track). Co-ordinated with hardware members and owned the computation subsystem.
3rd place (International). Solo, building PaperPlanes on a database I had never used before.
6th nationally, leading a team of six.
Track 1, containment standards. Our entry is Apocalypse.
Teaching math, English, and life skills to underprivileged children every week.
Ran logistics for ABV-IIITM Gwalior’s 1000+ attendee annual fest.
Editorial writing and storytelling for institute journalism.
Each project is written for a fast scan first, then a deeper read: problem, approach, system, stack, and proof. Every one links to its repository.
An AI agent holding a payment credential is one bad prompt away from an unbounded purchase. Access-token schemes can say an agent may spend, but not how much, on what, how many times, or for how long before that permission has to be re-earned.
Research assistants built as a vector store plus a chatbot cannot tell a superseded fact from a current one, and lose writes the moment more than one update lands at once.
Podman's May 2026 migration off Cirrus CI orphaned logformatter, the 38KB Perl script that classified every subtest as pass, fail, skip, or flake. Triage fell back to a human reading a bar graph, and the project does not re-run failing tests, so the obvious ground-truth signal (failed then passed on retry) is not handed to you.
Contested legal questions resist single consensus answers, and multi-agent systems steered by prompt instructions alone drift. Asked for 2 perspectives, the graph generated 6 to 8, a 3 to 4x overrun that exhausted Groq's token-per-minute ceiling before the debate resolved.
Team formation and capital discovery are the same failure wearing two costumes. Existing platforms either gate on post-traction signals a pre-seed founder cannot produce, or apply no verification at all and drown both sides in noise.
When autonomous agents buy from shared sellers, they collide: two agents bid against each other for the same product, agents waste rounds bidding under a seller's hidden floor, and agents buying different products from one seller lose the wholesale discount they jointly earned.
Most AI assistants answer the current prompt but lose long-term builder context, project momentum, and evolving preferences.
The July 2026 frontier-model sandbox escape had no public standard you could test a lab against. And the evidence is thin: there is no 17,600-action corpus, only 8 phase totals, 5 day counts, and 14 example commands in the public record.
Monitors in AI control are usually labelled as monitors. If an untrusted agent can tell it is being watched, it can choose what to show the watcher.
Water and walk reminders are notifications, and notifications get dismissed without being read.
Weighted toward what can actually be evaluated: the tools, the project choices, the tradeoffs taken, and the numbers behind each claim.
Call a real race condition an infra blip and you have told a maintainer to press re-run on a genuine bug. So abstention became a first-class verdict, and the harness reports abstention rate next to accuracy and counts the dangerous confusion separately. A classifier that says nothing is recoverable. One that is confidently wrong is not.
From Podman Flake AgentAsked politely for two perspectives, the graph produced six to eight, a 3 to 4x overrun that burned the token ceiling before the debate resolved. No amount of prompt rewording fixed it. Moving perspective count into a state schema the moderator schedules against fixed it completely, because the constraint stopped being something the model could ignore.
From RabbitHoleA fact is not static. It gets recorded, superseded, contradicted, and sometimes restored. Destructive updates make the current state cheap to read and the history impossible to recover. Closing rows with valid_to instead of overwriting them keeps a harder question answerable: what did we believe last month, and were we right?
From PaperPlanesI am open to AI engineering work where agentic workflows, retrieval, and product implementation meet. Every claim on this site traces to a case study and a repository, so the fastest way to evaluate me is to read one and then go check it.